The NEXT LAST R600 Rumours & Speculation Thread

[INKster]

So there is a difference in clocks between the XT & XTX:

XT : 750/830
XTX : 800/1100

Also now Fuad has changed his tune saying that XTX might not be delayed after all. :lol:

Hey, he did say "Q3".
Q3 begins on the 1st of July. There's still time, especially if the XT variant is slated to the second half of May.
We can't be sure yet that the XTX is anything less of a "phantom edition" than the rumored 8800 Ultra.

[Geeforcer]

We will probably have to wait a while to learn about ALU structure of R600. I wonder if use of "320" as opposed to the count of super-scalar ALUs we've seen in the past (remember the "48 now, up to 96 some time in the future") reflects underlying architectural differences from the past previous generations or is more of a marketing response to G80s scalar ALUs.

[caboosemoose]

Well, Dailytech doesn't sign NDA's.

Indeed. They let Anandtech do the NDA signing and then allow the holes in their non-existent chinese walls do their magic.

[INKster]

Indeed. They let Anandtech do the NDA signing and then allow the holes in their non-existent chinese walls do their magic.

I had no idea "Sven Olsen" was a chinese name. :wink:

[silent_guy]

Scalar architectures are usually more efficient than vector ones because they can work on more than one instruction (at different stages) per clock cycle per ALU.
No.
It doesn't matter if it's scalar or vector: in both cases, you will have as many instructions in flight as there are pipeline stages. The number of pipeline stages is orthogonal to the machine being scalar or vector.

As Frank wrote, scalar is more efficient because you won't 'waste' a part of your vector ALU when you're just executing scalar operations.

The greater amount of stages also has the interesting side-effect of boosting clockspeeds at the cost of latency,...
Yes., though the latency is not really a big deal here. A deeper pipeline also increases the chances of stalls due to instruction dependencies (which can be avoided on a GPU by scheduling a different thread.)

[INKster]

Yes., though the latency is not really a big deal here. A deeper pipeline also increases the chances of stalls due to instruction dependencies (which can be avoided on a GPU by scheduling a different thread.)

Can't those stalls be fought with greater L1/L2 cache sizes on-chip ?
We know G80 uses at least part of a unified L2 per 16 ALU's, but i'm not certain how much of it there is, and if it's optimized to such workloads.

The increasing use of both the R600 and the G80 as "GPGPU" processors may explain part of this strategy, because that kind of code is not exactly the same as that of graphical computations and may need more branching, therefore, more cache.

[Geeforcer]

The real question is, how smart/effective R600 scheduler is. If you have let's say 100 scalar operations, how will the ALUs be allocated?

[Razor1]

many vertex calculations are vector calculations maybe thats why the geometry shaders seem to be more potent because of better utilization?

[silent_guy]

Can't those stalls be fought with greater L1/L2 cache sizes on-chip ?
I'm talking about cases where instruction B depends on the result of A instruction, the previous one. If the multiplier is pipelined, then it takes a few cycles for the result of instruction A to complete. As a result, instruction B has to wait.

The longer your ALU pipeline, the higher the chance that this can happen, so your compiler has to be smart about it.

[INKster]

The longer your ALU pipeline, the higher the chance that this can happen, so your compiler has to be smart about it.

So, i'd take it that the G8x architecture still has more to gain by optimizing the compiler/driver software combination than R6xx, right upfront at least.
Am I understanding this correctly ?

[silent_guy]

many vertex calculations are vector calculations maybe thats why the geometry shaders seem to be more potent because of better utilization?
Aren't that vast majority of shader operations vector transformations, irrespective of them being fragment, vertex or geometry? If so, then that's why the difference between the two isn't that large. If you're going to do GPGPU operations that can't be vectorized, the difference in efficiency should be much larger.

That said, I've never written 3D shaders, so maybe somebody else can given an idea about how common it is to have scalar operations, for loops etc?

[silent_guy]

So, i'd take it that the G8x architecture still has more to gain by optimizing the compiler/driver software combination than R6xx, right upfront at least.
Am I understanding this correctly ?

I don't think it's going to be much a problem in practice: on a scalar machine, 3D vector operations have to be unrolled into 3 separate operations. Those will be executed after each other, will not be dependent, and won't stall. So that will avoid the problem in most cases right there.
Also, this kind of compiler instruction scheduling is a rather trivial algorithm in the bag of compiler writers.

[Jawed]

So, i'd take it that the G8x architecture still has more to gain by optimizing the compiler/driver software combination than R6xx, right upfront at least.
That's an unknown, since the architecture itself isn't known...

Jawed

[Mariner]

many vertex calculations are vector calculations maybe thats why the geometry shaders seem to be more potent because of better utilization?

Didn't the rumours say that G80 didn't really offer much in the way of support for Geometry Shaders? Or was this just a lack of support from early drivers? I can't really remember what the perceived wisdom was regarding GS support. Does anyone have any more information about this topic or have I just missed something very obvious somewhere?

If the rumour was true, it certainly wouldn't be the first time that NV have skimped on support for a feature in order to produce good performance in other areas. Dynamic branching performance, for example, in the last gen of products.

[INKster]

That's an unknown, since the architecture itself isn't known...

Jawed

DéjÃ* Vu ? (http://forum.beyond3d.com/showpost.php?p=975660&postcount=2738)

Don't tell me you're under NDA too. ;)

[Razor1]

Didn't the rumours say that G80 didn't really offer much in the way of support for Geometry Shaders? Or was this just a lack of support from early drivers? I can't really remember what the perceived wisdom was regarding GS support. Does anyone have any more information about this topic or have I just missed something very obvious somewhere?

If the rumour was true, it certainly wouldn't be the first time that NV have skimped on support for a feature in order to produce good performance in other areas. Dynamic branching performance, for example, in the last gen of products.


With unified shaders it seems unlikely they could skimp on GS.

[Razor1]

Aren't that vast majority of shader operations vector transformations, irrespective of them being fragment, vertex or geometry? If so, then that's why the difference between the two isn't that large. If you're going to do GPGPU operations that can't be vectorized, the difference in efficiency should be much larger.


That is true

[Jawed]

No I'm not under NDA. I just think it's unlikely that R6xx is a merely "vector ALU" architecture, certainly not in the sense of R5xx or Xenos.

Obviously, there's a possibility it is nothing more than vector ALU.

I think it's unlikely simply because the ALU utilisation problem is well-established. It's not something that would have bitten GPU designers only in the last year or two.

Jawed

[Frank]

No I'm not under NDA. I just think it's unlikely that R6xx is a merely "vector ALU" architecture, certainly not in the sense of R5xx or Xenos.

Obviously, there's a possibility it is nothing more than vector ALU.

I think it's unlikely simply because the ALU utilisation problem is well-established. It's not something that would have bitten GPU designers only in the last year or two.

Jawed
Yes, but remember the responses we got when discussing fine-grained sceduling for the R600. Sure, the G80 does most of that, but ATi was a firm believer of large batches and wide operations. Although they clearly wanted to be able to schedule things fast and more narrow, for better utilization, dynamic branching and external data fetching.

That sounds like an improved Xenos using 8*8 blocks of vector ALUs to me.

[Jawed]

Didn't the rumours say that G80 didn't really offer much in the way of support for Geometry Shaders? Or was this just a lack of support from early drivers? I can't really remember what the perceived wisdom was regarding GS support. Does anyone have any more information about this topic or have I just missed something very obvious somewhere?
It's extremely unlikely that G80 is functionally deficient. D3D10 doesn't provide any options for GS functionality. Obviously there's always the possibility of bugs...

GS performance is a function of more than just ALU capability, though. e.g. ability to read multiple input buffers, size of internal buffers, constraints on the number of primitives in-flight at any one time etc. Fuzzy stuff that we'll prolly never get any detailed insights on.

All we'll prolly get is game benchmarks or 3DMk synthetics.

There's a general clue that GS is "hard": some of the D3D10 specifications relating to GS were reigned-in because implementing this functionality is costly. The most obvious limitation is that only 1024x fp32s can be generated per input primitive. When lots of vertices are generated by GS, they have to be buffered somewhere before they're consumed by the setup/rasterisation stage. Considering that DX9 GPUs can only cache tens of vertices, the ability of the GS to potentially generate hundreds of extra vertices per input triangle puts quite a strain on the vertex cache.

Additionally, the streamout stage, which naturally follows on from GS (writing GS results to one or more vertex buffers), has constrained capabilities. Again, indicating that it's costly to implement.

Jawed

[Frank]

An architecture that is evolutionary rather than revolutionary always has more easy gains. If you come up with a totally different engine (like the G80), you have to swallow the numerous problems you're going to face trying to "emulate" the old behaviour. Such architectures only start to shine when they're used to the metal.

R600 is almost surely a better fit for DX10. But the G80 has a lot more potential.

[jimmyjames123]

Frank, when you say that G80 has a lot more potential, are you referring to having a lot more potential in DX10 applications compared to what it has right now?

[Jawed]

Yes, but remember the responses we got when discussing fine-grained sceduling for the R600.
I've got no current reason to dismiss those ideas. In fact the patent application I linked yesterday, for vertex calculation threading, appears to be conceptually identical to what we were discussing :mrgreen:

http://forum.beyond3d.com/showpost.php?p=975199&postcount=2645

That sounds like an improved Xenos using 8*8 blocks of vector ALUs to me.
I'm leaning towards R600 being a four-unit architecture, i.e. each shader unit is capable of 80 MADs, with, say, 4 TMUs and 4 ROPs. That's as far as I go...

Jawed

[jimmyjames123]

Didn't the rumours say that G80 didn't really offer much in the way of support for Geometry Shaders? Or was this just a lack of support from early drivers? I can't really remember what the perceived wisdom was regarding GS support. Does anyone have any more information about this topic or have I just missed something very obvious somewhere?

If the rumour was true, it certainly wouldn't be the first time that NV have skimped on support for a feature in order to produce good performance in other areas. Dynamic branching performance, for example, in the last gen of products.


In NV's tech documentation (http://www.nvidia.com/page/8800_tech_briefs.html) it says "New Features implemented in GeForce 8800 Series GPU's that work in concert with DirectX 10 features include geometry shader processing...". In fact, the document repeatedly talks about geometry shaders. It seems that this is something NV will pay big attention to moving forward with DX10 hardware, even if functionality/performance is not fully exposed yet.

[Love_In_Rio]

An architecture that is evolutionary rather than revolutionary always has more easy gains. If you come up with a totally different engine (like the G80), you have to swallow the numerous problems you're going to face trying to "emulate" the old behaviour. Such architectures only start to shine when they're used to the metal.

R600 is almost surely a better fit for DX10. But the G80 has a lot more potential.

Also had pentium 4 if it had reached 10 ghz as previously aimed, and i will agree with you if the performance per watt is better in G80 architechture. On the other hand, don´t you think it will help G80 greatly that geometry shaders hasn´t been finally punched with this direct x revision ?

[Frank]

Frank, when you say that G80 has a lot more potential, are you referring to having a lot more potential in DX10 applications compared to what it has right now?
Probably, when the drivers and the understanding of how to map DX10 to the G80 architecture matures. But DX10 is too restrictive. Look at CUDA, or the next implementation of the nVidia development kits for the big improvements. Which might be more suited to doing things like preprocessing of geometry, ray-tracing and physics than simply crunching out DX10 graphics faster.

I expect things like destructible terrain, cloth and other things that deform geometries and fluid mechanics to shine most.

[Frank]

I've got no current reason to dismiss those ideas. In fact the patent application I linked yesterday, for vertex calculation threading, appears to be conceptually identical to what we were discussing :mrgreen:

http://forum.beyond3d.com/showpost.php?p=975199&postcount=2645
Interesting! I didn't spot the connection when you posted it.

I'm leaning towards R600 being a four-unit architecture, i.e. each shader unit is capable of 80 MADs, with, say, 4 TMUs and 4 ROPs. That's as far as I go...

Jawed
I've become more cautious as well. ;)

[Geeforcer]

Correct me if I am wrong, but isn't entire outlook of G80/R600 relative GS perfromace based on one months-old one-liner blurb with very little context?

[Frank]

In NV's tech documentation (http://www.nvidia.com/page/8800_tech_briefs.html) it says "New Features implemented in GeForce 8800 Series GPU's that work in concert with DirectX 10 features include geometry shader processing...". In fact, the document repeatedly talks about geometry shaders. It seems that this is something NV will pay big attention to moving forward with DX10 hardware, even if functionality/performance is not fully exposed yet.
Definitely. Because they want you to use CUDA to unlock that potential. Most likely including the missing MUL.

And that would be a nice one-up against the competition at the same time.

[Ailuros]

Didn't the rumours say that G80 didn't really offer much in the way of support for Geometry Shaders? Or was this just a lack of support from early drivers? I can't really remember what the perceived wisdom was regarding GS support. Does anyone have any more information about this topic or have I just missed something very obvious somewhere?

If the rumour was true, it certainly wouldn't be the first time that NV have skimped on support for a feature in order to produce good performance in other areas. Dynamic branching performance, for example, in the last gen of products.

I thought it has been established by now that developers will get geometry shaders that don't suck with D3D10.1 (or beyond?). If the terms are too harsh for some then replace them with higher efficiency.

In any case as has been already noted what exactly do you think is the difference for a D3D10 compliant unified shader ALU, when it's called to handle pixel/vertex or geometry shaders?

Geometry shader support aren't optional for D3D10 compliance; what IHVs currently need is first better drivers and then a more flexible version of the same API that allows higher GS efficiency.

[Frank]

On the other hand, don´t you think it will help G80 greatly that geometry shaders hasn´t been finally punched with this direct x revision ?
Good question. Because it's non-standard, it will not be universally accepted. But it does supply a very nice angle to offer to Microsoft, when they're going to design DX11. Because the possible gains are great.

[Frank]

Correct me if I am wrong, but isn't entire outlook of G80/R600 relative GS perfromace based on one months-old one-liner blurb with very little context?
Not really, if you followed it, read between the lines and look closely at the patents. But it does belong in this thread: R600 Rumours & speculation.

[Ailuros]

"D3D11" might or should go a few steps further than just efficient geometry shaders. Topology might be nice, but fully programmable tesselation (think HOS f.e.) combined with it might be the next goal to reach.

[frameavenger]

"D3D11" might or should go a few steps further than just efficient geometry shaders. Topology might be nice, but fully programmable tesselation (think HOS f.e.) combined with it might be the next goal to reach.

But, Is R600 fully programmable tesselation a DX10 feature:?:

[Frank]

"D3D11" might or should go a few steps further than just efficient geometry shaders. Topology might be nice, but fully programmable tesselation (think HOS f.e.) combined with it might be the next goal to reach.
Definitely. But, why stop there? Probably the most major problem is still location. If they can implement a mechanism that will accelerate finding all the objects at a certain location, over time, they have solved half the developer nightmares.

Like automatic filtering on textures, it would be a huge gain if you could simply upload all your geometry to the GPU and have it determine what object/polygon/vertex collection is intersecting.

[leoneazzurro]

R600 is almost surely a better fit for DX10. But the G80 has a lot more potential.

I think also R600 has a lot of untapped potential at disposal. I.e. if the units have some degree of vector arrangement, it is possible that efficiency is -at the moment- lower than G80, so if there is a way to increase efficiency with driver releases and/or compiler optimization, we could see higher performances on R600 parts, too.

[Frank]

I think also R600 has a lot of untapped potential at disposal. I.e. if the units have some degree of vector arrangement, it is possible that efficiency is -at the moment- lower than G80, so if there is a way to increase efficiency with driver releases and/or compiler optimization, we could see higher performances on R600 parts, too.
Certainly. But, like nVidia has a hard time getting the G80 to perform well, simply because it is so very different, the possible gains for the G80 with a next 3D API (DX11/OpenGL2.x) are much bigger as well.

So, in the end it's mostly about politics: who can get the API writers to make the interface that benefits them most?

For DX9, it was clearly ATi. For DX11, I expect it to be nVidia. The roles are reversed.

Simply look at the performance, while considering that the G80 has the hard part of the current deal.

[leoneazzurro]

Certainly. But, like nVidia has a hard time getting the G80 to perform well, simply because it is so very different, the possible gains for the G80 with a next 3D API (DX11/OpenGL2.x) are much bigger as well.

So, in the end it's mostly about politics: who can get the API writers to make the interface that benefits them most?

For DX9, it was clearly ATi. For DX11, I expect it to be nVidia. The roles are reversed.

Simply look at the performance, while considering that the G80 has the hard part of the current deal.

I don't see why G80 should be gaining more than R600 with next 3D API or future releases of drivers (except for the missing MUL), considering that we don't know at the moment the R600 architecture. It may be or it may be not, we must see. And, for DX10, I don't think it was Nvidia that dictated the rules, considering also ATI ties with Microsoft.
As for the performances, we must still see which part really performs "better", (and, I should add, what "better" means in this case) and with more mature drivers from each part, and in DX10 tests. At the moment, we could agree that it's a very close race between NV and ATI parts, with the latter being really late.

[Silent_Buddha]

I think it's a bit premature to crown G80 as the most advanced most difficult to optimize for architechture until we've had a chance to look at R600 and see what ATI can and cannot do with it.

There's always the possibility that G80 is difficult to get to work at top theoretical speed due to some bug in the silicon. NV has had similar issues in the past with promised features that did not work correctly at launch and required a refresh to get working correctly. Video acceleration anyone?

Although if that is the case, I'd expect the refresh would end up a whole lot faster than G80.

Also, if politics has anything to do with this, then ATI has a much better chance of dictating the direction of DX11. Considering their currently ties to Microsoft and that MS still has an axe to grind with Nvidia in regards to the Xbox chip price restructing that Nvidia did as well as the licencing fees that Nvidia insisted on for ANY sort of backwards compatability in the Xbox 360. I'm sure MS is still not happy paying Nvidia royalties on every Xbox 360 sold due to having software emulation for Xbox compatability.

Regards,
SB

[Ailuros]

But, Is R600 fully programmable tesselation a DX10 feature:?:

R600 is supposed to have what exactly? :roll:

[Ailuros]

I don't see why G80 should be gaining more than R600 with next 3D API or future releases of drivers (except for the missing MUL), considering that we don't know at the moment the R600 architecture. It may be or it may be not, we must see. And, for DX10, I don't think it was Nvidia that dictated the rules, considering also ATI ties with Microsoft.

APIs get defined from Microsoft + all graphics IHVs. I've no idea though if small IHVs have a vote on that board, but I have an extreme hard time to believe that a Microsoft representative waltzes into the room like a dictator and forces down everyone's throat what is supposed to be of their best interest.

I've no single doubt that decisions are usually taken under perfectly democratic manners, which means that the majority wins.

As for the performances, we must still see which part really performs "better", (and, I should add, what "better" means in this case) and with more mature drivers from each part, and in DX10 tests. At the moment, we could agree that it's a very close race between NV and ATI parts, with the latter being really late.

Even if AMD should have a significant advantage under D3D10 (which I haven't seen as of yet any serious indications to support such a notion yet) it would affect upcoming games in what way exactly? Triple A titles like Crysis, UT3 (or whatever they call it nowadays) etc. are all true blue D3D9.0 games that will get D3D10 paths for some performance optimisations. Unless of course anyone expects G80s to not benefit from those performance optimisations at all. I personally don't think so.

[Frank]

I think it's a bit premature to crown G80 as the most advanced most difficult to optimize for architechture until we've had a chance to look at R600 and see what ATI can and cannot do with it.

There's always the possibility that G80 is difficult to get to work at top theoretical speed due to some bug in the silicon. NV has had similar issues in the past with promised features that did not work correctly at launch and required a refresh to get working correctly. Video acceleration anyone?

Although if that is the case, I'd expect the refresh would end up a whole lot faster than G80.

Also, if politics has anything to do with this, then ATI has a much better chance of dictating the direction of DX11. Considering their currently ties to Microsoft and that MS still has an axe to grind with Nvidia in regards to the Xbox chip price restructing that Nvidia did as well as the licencing fees that Nvidia insisted on for ANY sort of backwards compatability in the Xbox 360. I'm sure MS is still not happy paying Nvidia royalties on every Xbox 360 sold due to having software emulation for Xbox compatability.

Regards,
SB
Agreed. Interesting times. :D

[SugarCoat]

MS still has an axe to grind with Nvidia in regards to the Xbox chip price restructing


You mean when Microsoft dropped the price of the Xbox dramatically in the first year of introduction and insisted nVidia follow suite with their chip price even thought it violated contract? :roll:

I think you're really grasping with even mentioning that here. If anyone is going to directly impact what direction DirectX heads its going to be developers.

[Frank]

You mean when Microsoft dropped the price of the Xbox dramatically in the first year of introduction and insisted nVidia follow suite with their chip price even thought it violated contract? :roll:
As strange as it might sound, that is very much accepted business practice.

[Razor1]

Also developers influence the IHV's and the IHV's influence the API builders. I don't see why MS will alienate nV doesn't make much sense at this point (for something that happened years ago) Business and grudges really don't mix well specially if that grudge can hurt you in the long run. If more developers are using the g80 tech they will be asking for what ever weaknesses or possible strengths of that tech to be fixed,added,enhanced in the next Dx. Of course MS has thier own plans too, depending on which IHV has better implimentation vs. time or difficulty of implimentation probably will end up in the API as well. Add xbox 360 to the mix yeah that too will have influence on dx10.1 and up, but to a lesser extent since its a closed box.

[leoneazzurro]

APIs get defined from Microsoft + all graphics IHVs. I've no idea though if small IHVs have a vote on that board, but I have an extreme hard time to believe that a Microsoft representative waltzes into the room like a dictator and forces down everyone's throat what is supposed to be of their best interest.

I've no single doubt that decisions are usually taken under perfectly democratic manners, which means that the majority wins.

Majority in this case could only be MS+ATI or MS + Nvidia or all three together. I don't think neither ATI or Nvidia could press MS to do anything it does not want to do. I remember that there were indeed some issues in DX9 development that led also to the problems with Geforce FX. But I can be wrong.



Even if AMD should have a significant advantage under D3D10 (which I haven't seen as of yet any serious indications to support such a notion yet) it would affect upcoming games in what way exactly? Triple A titles like Crysis, UT3 (or whatever they call it nowadays) etc. are all true blue D3D9.0 games that will get D3D10 paths for some performance optimisations. Unless of course anyone expects G80s to not benefit from those performance optimisations at all. I personally don't think so.

I never said ATI will have a performance advantage. I only said it's premature to say G80 is better, and will perform much better in the future because it's more difficult to optimize. From my point of view, it will be more difficult to optimize dor R600 if it has non-scalar ALUs.
I wrote about DX10 because it's true that today's games are DX 9, but to see which card is more "future-proof" (as NV40 big selling point of supporting SM3.0 was some years ago) we need to know also which architecture will perform better not only in DX9, but also in DX10.
That's all.

[compres]

Majority in this case could only be MS+ATI or MS + Nvidia or all three together. I don't think neither ATI or Nvidia could press MS to do anything it does not want to do. I remember that there were indeed some issues in DX9 development that led also to the problems with Geforce FX. But I can be wrong.




I never said ATI will have a performance advantage. I only said it's premature to say G80 is better, and will perform much better in the future because it's more difficult to optimize. From my point of view, it will be more difficult to optimize dor R600 if it has non-scalar ALUs.
I wrote about DX10 because it's true that today's games are DX 9, but to see which card is more "future-proof" (as NV40 big selling point of supporting SM3.0 was some years ago) we need to know also which architecture will perform better not only in DX9, but also in DX10.
That's all.

Usually ATI gpus are more "future proof", meaning they perform better than their nvidia counterparts as time goes by. An extreme example is the x1800 vs the 7800, but there are other examples.

I would like someone to give me an example in recent history where nvidia has a card that later on gets even more better in relation to the competing ATI gpu, not because I want to take sides, but because I can't honestly remember.

That said I might not buy an ATI card again until they get their act together in Linux support.

[INKster]

Usually ATI gpus are more "future proof", meaning they perform better than their nvidia counterparts as time goes by. An extreme example is the x1800 vs the 7800, but there are other examples.

I would like someone to give me an example in recent history where nvidia has a card that later on gets even more better in relation to the competing ATI gpu, not because I want to take sides, but because I can't honestly remember.

That said I might not buy an ATI card again until they get their act together in Linux support.

6600 vs X700 (SM 3.0, Purevideo, etc).

[Ailuros]

Majority in this case could only be MS+ATI or MS + Nvidia or all three together. I don't think neither ATI or Nvidia could press MS to do anything it does not want to do. I remember that there were indeed some issues in DX9 development that led also to the problems with Geforce FX. But I can be wrong.

Ironically ATI wasn't the only IHV back then that had adopted FP24 internal precision for D3D9.0. The late XGi's and S3's architectures had FP24, I'd guess split precision for Imagination and a question mark for Intel's relevant IGPs. All that tells me is that someone (probably ATI) came up with a better overall idea for D3D9.0 and it received a wider adoption too. Unless you're trying to tell me that ATI or NVIDIA have a supposed stronger vote as Intel in the relevant board. I've no idea about all the details, but I'd expect a giant like Intel to have more saying that the majority would figure.

Read B3D's early "DX-Next" article and compare it to that D3D10 ended up with. Yes IHVs are actually protesting if they can't fit it all in hardware and yes Microsoft IMO follows those protests and drafts get adjusted according to what the majority wants or can realize in hardware within an estimated/projected transistor budget.


I never said ATI will have a performance advantage. I only said it's premature to say G80 is better, and will perform much better in the future because it's more difficult to optimize. From my point of view, it will be more difficult to optimize dor R600 if it has non-scalar ALUs.
I wrote about DX10 because it's true that today's games are DX 9, but to see which card is more "future-proof" (as NV40 big selling point of supporting SM3.0 was some years ago) we need to know also which architecture will perform better not only in DX9, but also in DX10.
That's all.


Depends what everyone means with "better" exactly. In any case G80 is probably NV's largest departure yet from their past architectures, but I have doubts it's actually the core of all driver/compiler development hurdles. I'd blame Vista and the resources consuming dual OS support need first before anything else and I don't see ATI having less problems with that one exactly either. At least all the early public statements from developers since now pointed in the direction that both IHVs Vista drivers need still a lot of work.

Now if one tries to dig deeper into the architectural internas of each of the two architectures, I as a layman would also guess that NV's decision to go for "pure" scalar ALUs (well always in the usual "GPU consensus") might be better suited for the future and might be a trend that might see broader adoption some time in the future. Yet that's a fine technicality for folks that have a far better understanding than us when it comes to hardware design decisions. As it stands what mostly matters for us laymen is what comes out at the other end.

That said all indications of the recent past and the most recent ones indicate IMHLO a neck to neck race between R600 and G80, each of the two with it's distinct advantages. It's high time that we see something different for D3D10, we've seen G80 and it's been analyzed to death.

[DemoCoder]

Hmm... So a card (GTS) with 64gb/s bandwidth, and 230-345 GFlops of shading power is hanging in the same ballpark with a card with 100+gb/s bandwidith, and 476-512 GFlops. On paper, 1.4-2x the flops, and 50% more bandwidth. By all rights, this card should be steam rolling over the GTS. And perhaps it will in bandwidth limited scenarios, but it looks like in heavy shader limited scenarios, it's not putting a beatdown on the GTS.

Perhaps we are seeing the first signs of inherent inefficiencies of a vec5 USA vs a scalar USA. Of course, 1 set of small benchmarks is too little to make such claims, so we will have to wait to see a much larger and more varied set of workloads.

However, this seems to suggest that if NVidia creates a refresh with say, 12-16 TCPs, they're going to clean up.

[IbaneZ]

http://www.dailytech.com/ATI+Radeon+HD+2900+XTX+Doomed+from+the+Start/article7052.htm

XTX OEM vs 8800 GTX.

Some weird results. :???:

[Evildeus]

http://www.dailytech.com/ATI+Radeon+HD+2900+XTX+Doomed+from+the+Start/article7052.htm

XTX OEM vs 8800 GTX.

Some weird results. :???:If true and correct, those are horrible benchmarks....

[Freak'n Big Panda]

http://www.dailytech.com/ATI+Radeon+...rticle7052.htm

XTX OEM vs 8800 GTX.

Some weird results.


Wow. I hope to god the R600 doesn't turn out to be another NV30. Does anybody have any idea as to why we're seeing such sub par numbers?

[BlizzardOne]

750 core on the XTX? I would have expected atleast another 50-75 on the core over the XT to justify the extra "X", rather than a paltry 5mhz...

[Freak'n Big Panda]

The numbers almost make sense (minus the oblivion bench where something obviously went wrong) if you take into account that the XTX is just 200MHZ away from the XT in memory speed. I doubt the games used more than the 512MB frame buffer provided so the small performance delta between the two seems reasonable. If these benchmarks are accurate though ATI is really in for a though time. Competitive pricing is really the only thing that is going to sell these boards.

[IbaneZ]

The GTX is overclocked though, 650/1000.

But I think I'll wait for more previews/reviews before I kill myself. ;)

[BlizzardOne]

The GTX also has a slight memory OC (1000mhz vs 900mhz default) in it's favour.

IbaneZ beat me :)

[Rangers]

Wow, it's really over for the former ATI with these benches.

It's bad enough they were constantly months late, but one expected them to at least provide a slightly higher performing part.

I'm guessing G80 is just a better architecture on all fronts. ATI didn't go the scalar route, they were simply lagging, while Nvidia was busy sticking a knife in them. These architectures are years and billions of dollars in the making, so it's unlikely ATI will ever recover now given there finances.

I honestly say that about wraps it up for ATI. I guess the fact they're AMD now may keep them alive a bit longer, but AMD is gone too most likely.

I hope all you guys constantly championing the demise of AMD on this forum, (and indeed, across the interwebs) and you know who you are, are happy when you pay Intel monopoly prices for your CPU's (and I suppose, Nvidia monopoly prices on your GPU's).

Fact is Intel just made some very nice price cuts that put Core2dou in my range, not to mention the excellent price cuts AMD just enacted. The fact is Intel would not have reduced without AMD, and that is our future.

Edit: I guess, in trying to look for excuses for ATI, I can come up with this: The G80 is Oc'd, and the XTX may be below release clocks (If Inq's 800mhz core on XTX figure was true). Adding 50mhz to the XTX gains ~7%, whil dropping the GTX to it's stock (really, the fair thing) drops it ~12%, you have ~20% right there, if you assume ATI can do some driver tweaking or something, maybe it could be competitive.

But really, that's just a bit of grasping..

[Geeforcer]

Humm, the negligible performance difference between the XT and XTX (assuming the tests are accurate, of course) would indicate that performance limitations for both parts lie somewhere else other then bandwidth - in fact it would seem that R600 is unable to take advantage of the enormous bandwidth supplied by 512bit bus, at least in the Dailytech's tests.

[nAo]

Wow, it's really over for the former ATI with these benches.
You must be kidding, it wouldn't be 'over' even if those benchmarks were correct.
Why don't you just wait real reviews with final hw and sw? C'mon..don't be to pessimistic :)

[Rangers]

You must be kidding, it wouldn't be 'over' even if those benchmarks were correct.
Why don't you just wait real reviews with final hw and sw? C'mon..don't be to pessimistic :)

It's over. Throw a party Mr nAo :grin:

I mean, I guess it's not really over..but it could be.

OK here's this, if Barcelona really kicks as much ass as Fudzilla's benchmarks say it does, then AMD seems to be on the ball. Maybe they can ride this ATI foible out with good mid/upper-range parts ($400 XT) and whip that company into shape in the meanwhile.

Again, grasping though..

Edit: I'm wondering if once again ATI didn't texture limit themselves ala the R520/80? Or if the efficiency of the scalar G80 is just that much higher? Because on the surface, the shader power should be comparable, even given some efficiency loss, with ATI claiming 320 stream processors. Sure Nvidia's 128 are double pumped, but they're still not even up to a comparable 256 scalars.

[Geeforcer]

You must be kidding, it wouldn't be 'over' even if those benchmarks were correct.
Why don't you just wait real reviews with final hw and sw? C'mon..don't be to pessimistic :)

Nonsense. Remember when Nvidia had a high-end part that was lagging competition by a factor of 2 wider delta then Dailytech results show and we never ever heard from then again after that?

[Geeforcer]

Sigh... time for bi-weekly "why ATI/AMD will be fine" list

1) The benchmarks are preliminary.
2) Even if they hold, this will still be much better then NV30 vs R300.
3) Even if those benchmarks are correct, what about other tests? AA/IQ? Other features? DX10?
4) High-end is what share of the market, exactly? With Nvidia's mainstream lineup firmly in "just Ok" territory, ATI would have to do precious little to match their performance - and 65nm process will help then in long term.
5) ATI is a part of a larger company. Something tells me AMD didn't buy them based on 2900 XTX performance.

[Evildeus]

Sigh... time for bi-weekly "why ATI/AMD will be fine" list

1) The benchmarks are preliminary.
Ok

2) Even if they hold, this will still be much better then NV30 vs R300.Yes

3) Even if those benchmarks are correct, what about other tests? AA/IQ? Other features? DX10?Those are IQ bench but, you are right we don't know the quality.

4) High-end is what share of the market, exactly? With Nvidia's mainstream lineup firmly in "just Ok" territory, ATI would have to do precious little to match their performance - and 65nm process will help then in long term.Sure, but what about mind share? That's an important part of high end marketing.
5) ATI is a part of a larger company. Something tells me AMD didn't buy them based on 2900 XTX performance.Hope not :lol:

[Geeforcer]

Humm, maybe Faud was on to something (http://www.fudzilla.com/index.php?option=com_content&task=view&id=661&Itemid=1)....

[neliz]

The benchmarks are skewed..

There is hardly a performance drop for the XT going from 16x12 to 19x12 while the rest of the tests show something around a 20% drop.
They didn't test CoH on the XT? Howcome the XT doesn't lose performance in ES:4 going from 12x12 to 16x12, but the xtx drops 15%?
Sure, the GTX runs at Ultra speeds.. but since when is ES:4 CPU limited at 16x12? Checking THG's vga chart, even the GTX drops about 50% moving from 10x7 to 16x12

[Geeforcer]

Howcome the XT doesn't lose performance in ES:4 going from 12x12 to 16x12, but the xtx drops 15%?
Sure, the GTX runs at Ultra speeds.. but since when is ES:4 CPU limited at 16x12?

According to Kris (http://www.dailytech.com/article.aspx?newsid=7052&commentid=132318&threshhold=1&red=3253#comments):

We had a copy error on the 1600x1200 Oblivion benchmark (it was the same as the 1280x1024 row). That is corrected now.

[leoneazzurro]

If core clock is almost the same between the XT and XTX, it is not strange to see similar results as IMHO they are not going to be bandwidth limited with a 512 bit bus.
I think however that ATI will try to push clocks in the 800 MHZ range for XTX, otherwise XTX has no reason to exist.

[Kaotik]

It just doesn't make any sense that XT would be in fact faster in some tests than XTX :???:

[leoneazzurro]

It just doesn't make any sense that XT would be in fact faster in some tests than XTX :???:

That could be due to early drivers or an early board. But yes, it is strange.

[Geeforcer]

It just doesn't make any sense that XT would be in fact faster in some tests than XTX :???:

How do GDDR3 and GDDR4 latencies compare?

[Pete]

That GTX might be the BFG Geo spotted (http://forum.beyond3d.com/showpost.php?p=975074&postcount=599).

Has anyone seen such a disparity b/w 3DM and game #s in recent competing architectures? (For the love of all that is holy, don't say that GPU.) It's odd that the XT/X hangs with the GTX in 3DM06 at a decidedly non-bandwidth-limited (considering the cards)12x10 no AA, yet falls so far behind in games. Also curious is that the XTX's 25% greater bandwidth contributes roughly squat for R600's framerates, at least at the piddling 4xAA settings we're given.

But what's holding the R600 back? Its FEAR #s are in R580+ territory (http://www.techreport.com/reviews/2007q1/geforce-8800gts-320mb/index.x?pg=5), despite the fact that it's got 2x the bandwidth, presumably 1.5x more MAD power, and inevitably improved and/or simply more texturing and ROP units. Don't tell me it's Z-limited (CoH and FEAR were benched with high and soft shadows, respectively--they demand Z samples out the wazzoo, right?).

More obviously, something's clearly wrong with the R600s when their framerates drop so much going from 16x12 to just 19x12. In fact, HL2:Ep1, the only game that's packing HDR on top of AA, is the only one that hints at G80 being bandwidth-limited at that last res bump, at least moreso than R600 (b/c, given the other benches, I can't fathom what other limitation G80 might have relative to R600).

I've attached some thought-provoking Wavey-niscent fillrate graphs (y: pixels/s, x: pixels), for ships and giggles.

[neliz]

How do GDDR3 and GDDR4 latencies compare?

gddr4 latency is higher..

On the signaling front, the new GDDR4 protocol now expand the chip I/O buffer up to 8-bit per two cycles, allowing for greater sustained bandwidth during the burst transmission, but at expense of significantly increased CAS latency(CL), determined mainly by the double reduced count of the address/command pins and half-clocked DRAM cells, compared to GDDR3. The amount of addressing pins was reduced to half of GDDR3 core, and were used for power and ground, which also increases latency.

But even at that.. it would mean a 1950XTX would be slower than a 1900XTX in a lot of benchmarks..

[Kaotik]

I'll throw in a wild rumor - XTX will not be released at all.
To honor Fudzilla, Inq and so on, I'll ad "according to sources close to ATI" to that

[Rangers]

That GTX might be the BFG Geo spotted (http://forum.beyond3d.com/showpost.php?p=975074&postcount=599).

Has anyone seen such a disparity b/w 3DM and game #s in recent competing architectures? (For the love of all that is holy, don't say that GPU.) It's odd that the XT/X hangs with the GTX in 3DM06 at a decidedly non-bandwidth-limited (considering the cards)12x10 no AA, yet falls so far behind in games. Also curious is that the XTX's 25% greater bandwidth contributes roughly squat for R600's framerates, at least at the piddling 4xAA settings we're given.

But what's holding the R600 back? Its FEAR #s are in R580+ territory (http://www.techreport.com/reviews/2007q1/geforce-8800gts-320mb/index.x?pg=5), despite the fact that it's got 2x the bandwidth, presumably 1.5x more MAD power, and inevitably improved and/or simply more texturing and ROP units. Don't tell me it's Z-limited (CoH and FEAR were benched with high and soft shadows, respectively--they demand Z samples out the wazzoo, right?).

More obviously, something's clearly wrong with the R600s when their framerates drop so much going from 16x12 to just 19x12. In fact, HL2:Ep1, the only game that's packing HDR on top of AA, is the only one that hints at G80 being bandwidth-limited at that last res bump, at least moreso than R600 (b/c, given the other benches, I can't fathom what other limitation G80 might have relative to R600).

I've attached some thought-provoking Wavey-niscent fillrate graphs (y: pixels/s, x: pixels), for ships and giggles.

I'm going with my guess detailed above, texture limited.

There was even some rumor that R600 only had 16 TMU's, just like R580..at the time I thought it absurd. But even if the part had a more reasonable say, 24, that still very well could be what holds it back. Doesn't 8800 GTX have at least 32?

[leoneazzurro]

gddr4 latency is higher..

Yes, but latency normally is not a big issue in GPUs, buit it could indeed be that current drivers/boards are not well tuned for GDDR4 If R600 has a programmable memory controller like R5X0, it could be that there could be issues with application tuning (as happened with Doom3/Quake4 and early X1900 drivers) and different memory types

[leoneazzurro]

I'll throw in a wild rumor - XTX will not be released at all.
To honor Fudzilla, Inq and so on, I'll ad "according to sources close to ATI" to that

I think so, if XTX is indeed 750 MHz core.

[nAo]

it's not just latency, it's also about the minimum amount of data you can transfer and how this impacts performance.
I think we already discussed about this several times.

[Geeforcer]

gddr4 latency is higher..



But even at that.. it would mean a 1950XTX would be slower than a 1900XTX in a lot of benchmarks..

1950 has15GB/s on 1900, plus higher core speed. It would take some very special test for any memory latency to overcome such a huge bandwidth gap. In case of 2900 XTX vs XT, if neither is bandwidth-limited (a real possibility due to 512-bit bus) then higher latency might actually show up in the results.

[erick]

How probable would you all consider the scenario that AMD deliberately leaked a low-performance driver revision to Dailytech or its source, so as to let them make complete asses out of themselves by publishing these numbers?

[neliz]

1950 has15GB/s on 1900, plus higher core speed. It would take some very special test for any memory latency to overcome such a huge bandwidth gap. In case of 2900 XTX vs XT, if neither is bandwidth-limited (a real possibility due to 512-bit bus) then higher latency might actually show up in the results.

Shame I can't find the b3d 1950 review anymore...
http://forum.beyond3d.com/showthread.php?t=33000
Bandwidth advantage is discussed there.

[Geeforcer]

How probable would you all consider the scenario that AMD deliberately leaked a low-performance driver revision to Dailytech or its source, so as to let them make complete asses out of themselves by publishing these numbers?

You mean besides the fact that AMD has nothing to gain by embarrassing Dailytech and potentially a lot to lose by creating a negative perception of their own product due to a fairly reliable news site publishing so-so numbers?

[Lux_]

How probable would you all consider the scenario that AMD deliberately leaked a low-performance driver revisionVery low - the driver is probably on the CD that comes with the card. And what's the point in hiding the performance?

[neliz]

How probable would you all consider the scenario that AMD deliberately leaked a low-performance driver revision to Dailytech or its source, so as to let them make complete asses out of themselves by publishing these numbers?

The driver is basically cat7.4 there's obviously space to release a newer driver for the reviews since people went home with the cards now and there's some time to do proper testing with a newer driver. since the 8.36 drivers are just two weeks old I doubt there will be much improvements in that time frame though.

[Geeforcer]

Shame I can't find the b3d 1950 review anymore...
http://forum.beyond3d.com/showthread.php?t=33000
Bandwidth advantage is discussed there.

Well, I went back and looked at Anand's review... I was actually rather surprised at how close the two performed despite a 22.5% bandwidth gap.

[Mcmlxxx IV]

A couple things do not make sense about their "XTX" test. Why would AMD put the expense of all that GDDR4 on the board for virtually no performance improvement? Would they really set the clocks of the highest end XTX at only 5 mhz higher on the core? I think their "OEM" version is not representative of the XTX. I am also doubting their XT scores, too. I won't say why, though.

[nAo]

Sorry to be a broken record but as we already discussed many times latency and bandwidth alone are not enough to analyze memory performance.
You can have amazing bandwidth and latency and still be forced to transfer so many bytes in the minimum transaction that most of them would be masked away anyway as you don't need to write/read them.

[erick]

A couple things do not make sense about their "XTX" test. Why would AMD put the expense of all that GDDR4 on the board for virtually no performance improvement? Would they really set the clocks of the highest end XTX at only 5 mhz higher on the core? I think their "OEM" version is not representative of the XTX. I am also doubting their XT scores, too. I won't say why, though.

NDA? ;)

[NocturnDragon]

RD790 just got PCI Sig certified..

http://www.pcisig.com/developers/compliance_program/integrators_list/pcie/

It looks more and more like a completely new family testbed

Just as 20 days ago ;)

http://forum.beyond3d.com/showthread.php?p=963057&highlight=RD790#post963057

[HAL]

http://www.fudzilla.com/index.php?option=com_content&task=view&id=703&Itemid=1


ATI officially confirmed


ATI said in front of 150+ journalists that R600XT won't be able to compete with Geforce 8800 GTX. The dream is dead and Radeon HD 2900 XT won't be able to catch up with a six month old Geforce 8800 GTX.

ATI said that R600XT, Radeon HD 2900XT is about the performance of 8800 GTS. Radeon HD 2900 XTX won't be launched at mid may and will be delayed to Q3 2007. We already wrote about this here.

The dream that AMD can do something and return as the high end market leader s gone. R600 rails against G80. Congratulations on Nvidia.

:wink:

[neliz]

The driver is basically cat7.4 there's obviously space to release a newer driver for the reviews since people went home with the cards now and there's some time to do proper testing with a newer driver. since the 8.36 drivers are just two weeks old I doubt there will be much improvements in that time frame though.

It's terrible to quote yourself.
http://www.fudzilla.com/index.php?option=com_content&task=view&id=702&Itemid=1

Cat7.5 at launch, with working audio
8.37 is the codename

Radeon HD 2000 generation will get a new driver. Despite that everyone believes that the 8.361 will be the final one there will be one more scheduled for May.

The new driver is in release candidate phase today and should be ready with its WHQL signature in May.

ATI calls this piece of software 8.37 and this driver will be the first one to bring the audio capabilities in life. This driver is planed for both Vista and Windows XP.

We are not sure when will ATI get a Linux driver for this card.

[Kaotik]

If Fuad is on to something, so to say, with that "XTX in Q3", it will be 65nm R650 and probably named 2950 instead of 2900.

[Rangers]

I dont know guys, these numbers dont make sense all the time.

Example, the XT scored nearly, if not, as well as the 8800GTX in 3Dmark according to dailytech earlier..

Yet it falls behind by huge amounts in games?

[neliz]

I dont know guys, these numbers dont make sense all the time.

Example, the XT scored nearly, if not, as well as the 8800GTX in 3Dmark according to dailytech earlier..

Yet it falls behind by huge amounts in games?

If I read correctly, the whole ES:4 benchmark test consists of 5 seconds (if not less) of the moment you exit the dungeon and walk out in the open.

[satein]

I would like to suggest you guys to wait for a moment, or a couple of weeks. Since when ATi/AMD told us that the R6xx will target at DX9 gaming system? Is it, R600, supposed to do the job on DX10 and to run DX9, it would need some magic in the driver in order to get its performance out.

If R600 is 4-groups of 16 Vec5, how could it will be great for a fix ratio of the DX9 hardware??? [This according to Jawed posted some pages before]

Let us see till the end when we see how they get the way to bench DX10 performance. To me, G80 would be king for the DX9 gaming atm, but it might not imply that G80 will be the same king on DX10 gaming and other tasks.

Best regards,

[Evildeus]

I would like to suggest you guys to wait for a moment, or a couple of weeks.

If we were to wait, this wouldn't be a " The NEXT LAST R600 Rumours & Speculation Thread" ;)

[neliz]

I would like to suggest you guys to wait for a moment, or a couple of weeks. Since when ATi/AMD told us that the R6xx will target at DX9 gaming system? Is it, R600, supposed to do the job on DX10 and to run DX9, it would need some magic in the driver in order to get its performance out.

We have some words that it would be the fastest dx9 card available.

[CJ]

We have some words that it would be the fastest dx9 card available.



Didn't they say it would be 'their' fastest DX9 card available?

[neliz]

Didn't they say it would be 'their' fastest DX9 card available?

Yeah...

“The R600 will be [absolutely] the fastest DirectX 9 chip that we had ever built,� said Richard Huddy, the head of ATI Technologies’ software developers relations department, at a press conference in London, UK.

Mr. Huddy said that Xbox 360 game console, which sports developed by ATI Xenos graphics core with unified shader architecture and 48 shader processors, loses 20% to 25% performance in pixel-shader limited games, when its graphics chip is configured as non-unified, e.g.,16 processors work strictly on vertex shaders, whereas 32 are assigned for pixel shaders.


http://forum.beyond3d.com/showthread.php?t=31049

[Evildeus]

Yeah...
Not too great :cry:

[chavvdarrr]

5) ATI is a part of a larger company. Something tells me AMD didn't buy them based on 2900 XTX performance.
yep. If AMD had known R600 performance they wouldn't buy ATi.
Me thinks ATi knew in what sh*t they are, so they looked around for a helping hand ... and eventually instead of being dragged away from the swamp they'll drag AMD into it :evil:

To Sum it:

To Kaotik

512 bus vs 384
320 pus vs 128
80nm vs 90nm
Yet slower.

And yes, I count processing units, I don't care at what speed are they running!
NV are able to run on higher freqs, ATi are not, so, who made better decisions?
Right now, R600 looks like pure brute-force decision.

[erick]

yep. If AMD had known R600 performance they wouldn't buy ATi.
Me thinks ATi knew in what sh*t they are, so they looked around for a helping hand ... and eventually instead of being dragged away from the swamp they'll drag AMD into it :evil:

I think that the performance of R600 was the last thing on AMDs mind when they bought ATI. If you want a clear reason for this move, look to Fusion.

[trinibwoy]

If AMD and ATi didn't know R600 performance at the time of the acquisition then something is terribly wrong. They didn't buy ATI for R600.

[chavvdarrr]

If AMD and ATi didn't know R600 performance at the time of the acquisition then something is terribly wrong. They didn't buy ATI for R600.
fact is, making more expensive chip on more expensive PCB, 6 months later , ATi is unable to be faster than G80
They could have a winner with medium 256-bit chip, yet it seems they'll fail there too - availability in June and no 256b according to rumours.

Oh, and posponing R600 now seems reasonable - in the beginning they believed XTX will be on par with GTX so started producing expensive 1GB boards... and then suddenly someone compared speed and production costs... maybe I'm a bit paranoid.. .but frankly, last 1-2y both companies are performing from bad to worst :(

[Bjorn]

To Sum it:

To Kaotik

512 bus vs 384
320 pus vs 128
80nm vs 90nm
Yet slower.


It has been mentioned here a couple of times (by Chalnoth f.e) that the decoupled shader units of the G80 would be the biggest deciding factor between the R600 and G80.

Certainly seems that way now and definitely not in AMD's favour.


Oh, and posponing R600 now seems reasonable - in the beginning they believed XTX will be on par with GTX so started producing expensive 1GB boards... and then suddenly someone compared speed and production costs... maybe I'm a bit paranoid.. .but frankly, last 1-2y both companies are performing from bad to worst

I'm starting to think that the G80 was quite a nasty surprise for AMD.

[nAo]

I'm lost.. "decoupled shader units" what?

[Rangers]

Working out the theoretical shader power, ATI claims something like 320 stream procs (scalar ALU's) at ~750mhz. Nvidia has 128, of course they are "double" pumped. But they are at 1350 mhz..so Nvidia it's the same as if they had 256 at 675mhz.

Theoretically:

X2900XT=320@750
8800GTX=256@675

There must be basically one of two things going on imo, Nvidia is getting much more effiency out of their setup, or ATI is once again severly texture limited.

I'd strongly suspect the latter. I'm very curious what sort of texture specs the card has now...I suspect Jawed's rumur of 16 TMU's in R600 was in fact correct.

Come on guys, all the great minds here that know a million more times about GPU's than me and I'm on the forefront of figuring this out :lol? (unless Im completely wrong :)

I'm starting to think that the G80 was quite a nasty surprise for AMD.

I dont know why it would be. It was the standard 2x improvement on previous generation high that has always been the norm in GPU's. It's ATI that put out a deficient part (lots of benchmark cases people finding where R600 is apparantly not faster at all than X1950XTX if Dailytech benches are right), not Nvidia a great one.

[Panajev2001a]

...

[nicolasb]

You cannot justify buying out ATI for the chipset sector alone...You can if you're a very large CPU manufacturer looking to compete with an even larger CPU manufacturer whose principal advantage over you is high-quality motherboard chipsets and cheap integrated graphics solutions. Ever since the merger was announced I've been speculating that a post-merger ATI would be exiting the high-end graphics market completely. The latest R600 benchmarks have reinforced my opinion. :(

[Kaotik]

And yes, I count processing units, I don't care at what speed are they running!
NV are able to run on higher freqs, ATi are not, so, who made better decisions?
Right now, R600 looks like pure brute-force decision.
It's not just about the speed they're running at, it's also about what they can do (G80 stream processors can dual-issue MADD & MUL if my memory serves, but we don't (yet) know what the R600 stream processors can do).
I wasn't trying to argue which was better solution, just that they're different and shouldn't be compared "by the number of units"

[Bjorn]

I'm lost.. "decoupled shader units" what?

Sorry, just talking about the different clock domain.

[seahawk]

It is a rather simpel thing :

ATi has 150% more Shader units running at roughly 60% the speed. So to make up for the speed difference they should need roughly 210 units to compete with NV. then there is the question of efficency how flexible is a ATI VEC5 unit ?

VEC4+1 is a given. But whalelse can it do ?

[Lux_]

I suspect Jawed's rumur of 16 TMU's in R600 was in fact correct.If it's really the case, then ATI devrel has been living under huge pile of rock...

I wonder if ATI provides journalists with a tool to benchmark the cards in folding - AFAIK there is no better way to show how much more powerful the new generation is.

[CarstenS]

I wonder if ATI provides journalists with a tool to benchmark the cards in folding - AFAIK there is no better way to show how much more powerful the new generation is.
Yeah, especially since the F@H-client doesn't run on Nvidia-cards (yet?).

[IbaneZ]

http://www.fudzilla.com/index.php?option=com_content&task=view&id=703&Itemid=1


ATI officially confirmed


ATI said in front of 150+ journalists that R600XT won't be able to compete with Geforce 8800 GTX. The dream is dead and Radeon HD 2900 XT won't be able to catch up with a six month old Geforce 8800 GTX.

ATI said that R600XT, Radeon HD 2900XT is about the performance of 8800 GTS. Radeon HD 2900 XTX won't be launched at mid may and will be delayed to Q3 2007. We already wrote about this here.

The dream that AMD can do something and return as the high end market leader s gone. R600 rails against G80. Congratulations on Nvidia.

:wink:

Well, if that's the case NV might as well cancel the 8800 Ultra.

So what the hell happened? With the monster specs R600 has it should blow G80 away.

I still don't believe DT's benchies. :razz:

[vertex_shader]

Well, if that's the case NV might as well cancel the 8800 Ultra.

So what the hell happened? With the monster specs R600 has it should blow G80 away.

I still don't believe DT's benchies. :razz:

FUD campaign against ati and r600? :wink:

[Rangers]

Well, if that's the case NV might as well cancel the 8800 Ultra.


Ultra core clock is only 650 I believe.

Mem clock may be more, but it's no big deal performacewise really..

[Kaotik]

Ultra core clock is only 650 I believe.

Mem clock may be more, but it's no big deal performacewise really..

According to Dell's specs, 650MHz core 1080MHz mem.

[Rangers]

According to Dell's specs, 650MHz core 1080MHz mem.

I know, ironically same core clocks as the OC'd 8800 Dailytech used.

You might even claim they used an Ultra in a way..

[Kaotik]

I know, ironically same core clocks as the OC'd 8800 Dailytech used.

You might even claim they used an Ultra in a way..

DT used 650/1000

[leoneazzurro]

It's not just about the speed they're running at, it's also about what they can do (G80 stream processors can dual-issue MADD & MUL if my memory serves, but we don't (yet) know what the R600 stream processors can do).
I wasn't trying to argue which was better solution, just that they're different and shouldn't be compared "by the number of units"

yes, we do. In their teraflop-in-a-box presentation, AMD stated that they had 320 Multiply-accumulate units. That is, MADD units.
hence, 2 Flops per unit per clock cycle at best, versus 3 flops (counting the other MUL) for G80 units.

[Razor1]

I dont know guys, these numbers dont make sense all the time.

Example, the XT scored nearly, if not, as well as the 8800GTX in 3Dmark according to dailytech earlier..

Yet it falls behind by huge amounts in games?


hmm bandwidth really played an important roll here, definitly "unprecedented" :razz:

[Rangers]

yes, we do. In their teraflop-in-a-box presentation, AMD stated that they had 320 Multiply-accumulate units. That is, MADD units.
hence, 2 Flops per unit per clock cycle at best, versus 3 flops (counting the other MUL) for G80 units.

But yet one of the G80 Mul's supposedly is missing to date..so THATS not it...

I'll pretty much bet anyone here we're looking at 16 TMU's on R600..

[Geo]

Wow, it's really over for the former ATI with these benches.

I honestly say that about wraps it up for ATI. I guess the fact they're AMD now may keep them alive a bit longer, but AMD is gone too most likely.


Yeah, that's why NVIDIA went out of business in 2003. I sure miss them, don't you? No, wait, it was 2004 they went out of business when they had their second generation in a row of not being the performance leader. No, wait. . . .

[Jawed]

I feel particularly stupid giving Anand/Dailytech benchmarking credence, but what the heck.

There must be basically one of two things going on imo, Nvidia is getting much more effiency out of their setup, or ATI is once again severly texture limited.
If R600 is 16 TMUs, are G80's 64 TFs more efficient?

Was R600 designed to clock at 1GHz, but has fallen to 80% of that? (Seems unlikely, normally cards get released at ~ the target clocks - and this has had plenty of time to get there.)

I'd strongly suspect the latter. I'm very curious what sort of texture specs the card has now...I suspect Jawed's rumur of 16 TMU's in R600 was in fact correct.
That's a guess based on guessing that RV610 is 4 TMUs (could it be less?!), and is one-quarter the pipeline configuration of R600.

Come on guys, all the great minds here that know a million more times about GPU's than me and I'm on the forefront of figuring this out :lol? (unless Im completely wrong :)
Maybe this is like 8600GTS. Some think it's horrible, some don't? One thing I know it isn't is GFLOPs.

I dont know why it would be. It was the standard 2x improvement on previous generation high that has always been the norm in GPU's. It's ATI that put out a deficient part (lots of benchmark cases people finding where R600 is apparantly not faster at all than X1950XTX if Dailytech benches are right), not Nvidia a great one.
Yeah, apparently R600 is the same speed as X1950XTX :???:

Maybe ATI got distracted by all the gubbins they could throw in to making D3D10 a shiny new universe and forgot about performance?

Jawed

[Galduta]

In some the tests of Dailytech, the 1950xt surpasses the r600 , or the avantaje not is great.



Company of heroes 1280*1024:
1950 XTX 99 (tweaktown) E6600
2900 XTX 97 (dailytech) QX6800

Company of heroes 1600*1200:
1950 XTX 70 (tweaktown) E6600
2900 XTX 73 (dailytech) QX6800

http://www.tweaktown.com/articles/1084/6/page_6_benchmarks_company_of_heroes/index.html


FEAR 1280*1024:
1950 XTX 80 (tomshardware) No softshadow 4AA 8AF , X6800
2900 XTX 84 (dailytech) with softshadow 4AA 16AF , QX6800

FEAR 1600*1200:
1950 XTX 57 (tomshardware) No softshadow 4AA 8AF , X6800
2900 XTX 58 (dailytech) with softshadow 4AA 16AF , QX6800

http://www.tomshardware.com/2007/02/12/the_amd_squ... (http://www.tomshardware.com/2007/02/12/the_amd_squeeze/page6.html)

Humm and the 1950 xtx with SS+AA ? and the 1950 xtx at 750/ 2500 mhz with SS +AA in FEAR ?

[INKster]

Yeah, that's why NVIDIA went out of business in 2003. I sure miss them, don't you? No, wait, it was 2004 they went out of business when they had their second generation in a row of not being the performance leader. No, wait. . . .

Geo, you've tricked me. ;)
Back when i was considering a 8800 GTS vs a cheaper "R600 XL", you told me both would be good choices.
But now i can only imagine if ATI had dared to launch an even slower version of R600 to compete in price with the GTS, the performance would certainly be disastrous.

Having said that, the Geforce FX 5200 actually sold pretty well during those dark times, i think.

[Geo]

I know, ironically same core clocks as the OC'd 8800 Dailytech used.

You might even claim they used an Ultra in a way..


If that's a thousand dollar card (and it roughly is at Newegg right now), then you'd have to say "boutique" and an unfair comparison.

But then some of those numbers if extrapolated down to even reference G80 look like 8800 still wins. But that's still not enough numbers and scenarios to make me comfortable we can slot this card yet.

But if we assume, for the sake of argument, that Nvidia is going to keep the performance crown this generation, I don't see why that has to be a disaster for AMD/ATI any more than not having the performance crown was a disaster for Nvidia in the GF6 generation. It's surely not a GOOD thing for AMD, but it doesn't have to be a disaster. And, interestingly enough, it looks like some similar elements may come into play. A price/performance leader in the one down slot in 2900 XT for enthusiasts (think 6800GT), and possibly a very competitive midrange. At least there's no evidence yet that NV slammed the door on midrange, tho we'll see.

[Jawed]

But now i can only imagine if ATI had dared to launch an even slower version of R600 to compete in price with the GTS, the performance would certainly be disastrous.
Supposedly R600 is so slow that it only just outperforms 8600GTS. Which would make sense, they both have 16 TMUs and R600 is clocked a bit higher, right?...

Jawed

[Geo]

Geo, you've tricked me. ;)
Back when i was considering a 8800 GTS vs a cheaper "R600 XL", you told me both would be good choices.
But now i can only imagine if ATI had dared to launch an even slower version of R600 to compete in price with the GTS, the performance would certainly be disastrous.



Hmm? The reports I've seen suggest XT is being positioned against GTS price-wise, at better performance. So take your pick --better performance at the same price point you were considering previously, or a lower-priced GTS when NV makes the inevitable adjustments to their lineup to respond.

It still appears to me that NV has a pretty dang big hole in their lineup between GTS and GTX.

[INKster]

Hmm? The reports I've seen suggest XT is being positioned against GTS price-wise, at better performance. So take your pick --better performance or a lower-priced GTS when NV makes the inevitable adjustments to their lineup to respond.

It still appears to me that NV has a pretty dang big hole in their lineup between GTS and GTX.

That hole, is it price or performance-wise ?
I don't believe for a minute that Nvidia will not take advantage of the R6xx perceived weakeness (at this point anyway), and lower a GTX performance product down to the upper-level GTS 640MB price segment.
The Ultra could then pick up the 550+ dollar and up segment.

I mean, the QuadroFX 4600 clearly shows a technical tendency here.

[Rangers]

Supposedly R600 is so slow that it only just outperforms 8600GTS. Which would make sense, they both have 16 TMUs and R600 is clocked a bit higher, right?...

Jawed

So you're confirming 16 TMU's?

[CarstenS]

Would it be a great surprise, if AMD was going common sense and realized that with GDDR4 and immense clockrates those cards would be to expensive for "normal" buyers and cancelled the high-end models, reworked the top-class to accomodate the more affordable GDDR3 RAM and a silent dual-slot-cooling solution? :)

So they can sell enough cards to the segments that really matter. After all, Athlon FX and Pentium/C2D Extreme Editions are for a very very small market also.

edit:
Reason for this: Look at Kombatants citation floating around in one of the signatures here: A pleasant surprise to many people (with pleasant and many in bold). Additionally, Vijay mentioned in our interview not something about a performance hammer or G80-Killer but rather stressed the point, that R6x-Familiy will deliver a very good price-performance-ratio - which is usually not found in the high-end.

According to rumor mill, my (then) guess seemed not to far off the mark. ;)

[Geo]

That hole, is it price or performance-wise ?
I don't believe for a minute that Nvidia will not take advantage of the R6xx perceived weakeness (at this point anyway), and lower a GTX performance product down to the upper-level GTS 640MB price segment.

Then your complaint would be what, exactly? That you're getting more card for less money and you hate when that happens?

[Evildeus]

Is the 999$ price tag a given or not? Because othewise the MSRP of the G80 could be:
-599$ G80 ultra
-499$ G80 GTX
-399$ G80 GTS 640
-349$ G80 GTS 320

[neliz]

Is the 999$ price tag a given or not? Because othewise the MSRP of the G80 could be:
-599$ G80 ultra
-499$ G80 GTX
-399$ G80 GTS 640
-349$ G80 GTS 320

CJ heard it from multiple sources (for quite some time) and Newegg is quite an indicator as well (selling the bfg oc card at almost the same clockrates)

[LunchBox]

looking at the the benchmarks dome by DT...

it make me wonder if ATI's brilliant engineers were spread to thin coz of handling 2 of the two next-gen consoles' GPU...

[jamis]

Hmm..Only 16 TMU's would sort of explain the "R600 is much better in geometry shading" rumour, wouldn't it? I mean if it has more math power but it's heavily bottlenecked at texturing. But what's the point in that sort of part?:lol:

[INKster]

Then your complaint would be what, exactly? That you're getting more card for less money and you hate when that happens?

No, i'm complaining that a R600 XT purchase will likely force me to buy another PSU (as my 550W, dual-rail, 34amp one might not be enough).
And then there goes the price-advantage against the GTS 640MB...

[Jawed]

So you're confirming 16 TMU's?
No. But it seems a reasonable guess. Also, we don't know what the capabilities per TMU are. It's hard to be sure which texturing related patent applications relate to R5xx and which to R6xx, but I see signs that there's a new design for R6xx texturing.

Also, bear in mind that "texturing" under D3D10 is a really loose concept. Heretofore texturing has been a catch-all way of feeding sizable amounts of both variable and static data into the pixel shading pipeline (or vertex shading pipeline).

Under D3D10, these classes of variable and static data have been segregated cleanly based upon access patterns and data-structure complexity. Hence you have:

boring old mipmapped textures
mipmapped texture arrays (to solve the texture atlas problem)
texture buffers (non-mipmapped, 2D or 3D blobs of single-data points) good for deferred shading or physics or some other kind of intermediate data storage
constant buffers containing more odd-ball datastructures, e.g. 1001 different constants to describe all the different kinds of clothes worn by characters in a game (i.e. determining the lighting effects)(Vertex buffers are the other major type of buffer, arguably not much changed from DX9, but with better instancing support - I'm a bit fuzzy on this...)

The upshot is that what was once a TMU may now either be multi-functional (and therefore texture filtering is less important) or have siblings that are moderately similar in functionality but are optimised for the kinds of data involved.

We know G80 does all this D3D10 buffer stuff, but very little mention is made of the architectural details that make it happen...

Jawed

[CJ]

No, i'm complaining that a R600 XT purchase will likely force me to buy another PSU (as my 550W, dual-rail, 34amp one might not be enough).
And then there goes the price-advantage against the GTS 640MB...

Recommended PSU for the HD2900XT is 500W. For CrossFire AMD recommends a 750W PSU. So you should be fine with your 550W PSU.

[INKster]

Recommended PSU for the HD2900XT is 500W. For CrossFire AMD recommends a 750W PSU. So you should be fine with your 550W PSU.

We all know Wattage is just part of the answer.
Nvidia recommends 26amps for a GTS, 30amps for a GTX.

How much for this R600 XT in that respect ?

[nAo]

it make me wonder if ATI's brilliant engineers were spread to thin coz of handling 2 of the two next-gen consoles' GPU...
Make that one next gen console, Wii's GPU is not really a new design

[chavvdarrr]

Yeah, that's why NVIDIA went out of business in 2003. I sure miss them, don't you? No, wait, it was 2004 they went out of business when they had their second generation in a row of not being the performance leader. No, wait. . . .Question is:
Can AMD fight against 2 dominating players - Intel and NVidia ...

[nAo]

Question is:
Can AMD fight against 2 dominating players - Intel and NVidia ...
Well..ATI could 'fight' against NVidia very well, I don't see why suddenly they should not be able anymore (well, apart from stupid management decisions..)

[Evildeus]

Well..ATI could 'fight' against NVidia very well, I don't see why suddenly they should not be able anymore (well, apart from stupid management decisions..)More ressources allocation issues perhaps?

[Razor1]

yeah they just have to price their products accordingly, shouldn't be a problem, thats what AMD is doing anyways, only problem is they are bigger now, so they do have to cut down expenditure and reorganize, which they are dong that too.

[Anon Lamer]

If R600 really only have 16 TMUs then I guess its intended to run with HDR/high AA always on, otherwise I see no use for all that bandwith. If it can do that and lose less fps than G80 then I would say its not a dud after all.

[Geo]

If R600 really only have 16 TMUs then I guess its intended to run with HDR/high AA always on, otherwise I see no use for all that bandwith. If it can do that and lose less fps than G80 then I would say its not a dud after all.

Yeah, we haven't seen that scenario much yet I think in these leaked benchies. But it does look like that seems to be the thinking from what's leaked so far. Maximise the per-pixel resources rather than increasing the number of pixels. Which is really pretty much a continuation of what they were saying in the R580 interview. . .

[leoneazzurro]

But yet one of the G80 Mul's supposedly is missing to date..so THATS not it...

I'll pretty much bet anyone here we're looking at 16 TMU's on R600..

I don't remember when, but I read somewhere that the "missing" MUL recently is not really "so missing". Anyway, we should know if really G80's units are doing more work for clock cycle.

[trinibwoy]

I don't see why that has to be a disaster for AMD/ATI any more than not having the performance crown was a disaster for Nvidia in the GF6 generation.

It won't be a disaster regardless but it's a stretch to compare this to the GF6/X800 days if ATI has no competition at all for Nvidia's top two cards. Losing by a couple percent at the high-end is one thing - having to sell your "flagship" for $400 just to compete is a bit more serious I would think.

[Arnold Beckenbauer]

I don't remember when, but I read somewhere that the "missing" MUL recently is not really "so missing". Anyway, we should know if really G80's units are doing more work for clock cycle.

Here?
General Shading Performance (http://www.beyond3d.com/content/reviews/7/17)
:wink:

[Geo]

It won't be a disaster regardless but it's a stretch to compare this to the GF6/X800 days if ATI has no competition at all for Nvidia's top two cards. Losing by a couple percent at the high-end is one thing - having to sell your "flagship" for $400 just to compete is a bit more serious I would think.

Let's see what that top card is and what it costs. If the BFG is an indication, I don't see that as anything other than a preener. If they want to make a new single-card market at $900+, then g'bless and all that.

[Jawed]

If R600 really only have 16 TMUs then I guess its intended to run with HDR/high AA always on, otherwise I see no use for all that bandwith. If it can do that and lose less fps than G80 then I would say its not a dud after all.
Until we know the architecture for all kinds of buffer fetches (vertex, texture, constant) it's hard to judge how bandwidth-demanding the generic concept of "texturing" is. As I said earlier, D3D10's approach prolly means that you can't judge capability from the number of TMUs.

As much as anything, R600 might be more focussed on D3D10. Wasn't R300 seen as nothing more than a "slightly faster DX8", with AF, AA and SM2 taking centre stage? How long did ATI have to wait for that to pay-off? :lol:

Having said all this, 16 TMUs at anything like the same "rate" as R5xx seems rather stingy. 120GB/s+ for DX9 apps really is disproportionate if the TMUs aren't up to it - and CrossFire R5xx shows that there's plenty more performance to come in apparently texturing-limited games. Let alone G8x, which one way or another, seems to show the same.

16 TMUs, considered on their own, would make R600 a 4:1 ALU:TEX architecture. I've long thought that R5xx would prolly have performed the same if it was only 2:1 instead of 3:1. For DX9 apps, 4:1 may well just be beyond the pale.

But under D3D10, with the variety of buffers seen by the programmer, the "ALU:TEX" ratio may fall, it might be 2:1, if there are pipes dedicated to point sampling or constant buffer fetches. Xenos has 16 dedicated point samplers (which are mostly used for vertex buffer fetches) but they can also be used to perform non-filtered "texture fetches" or constant fetches (guessing about the latter).

Might be better to think of ALU:Fetch ratio.

Jawed

[Sound_Card]

I smell stink all over DailyTechs performance figures.

They place the HD 2900XTX as the same exact speed(actully a little slower than the R580+) as the x1950xtx. If that does not say enough, I don't know what else will will.

http://rage3d.com/board/showpost.php?p=1334912940&postcount=5212

Here is another thing I found wrong with dailytechs figures...

On 1280x1024 they tested the HD 2900XT getting 47.9fps in Oblivion. However, their latest benchmarks put the HD 2900XT getting 73.4fps in a resoluton of 1920x1200. Same CPU and same game settings.

[leoneazzurro]

Here?
General Shading Performance (http://www.beyond3d.com/content/reviews/7/17)
:wink:

Yes, thanks! But not only there. ;)

[trinibwoy]

Let's see what that top card is and what it costs. If the BFG is an indication, I don't see that as anything other than a preener. If they want to make a new single-card market at $900+, then g'bless and all that.

I'm as skeptical as the next guy with respect to these DT benchmarks - the XT isn't living up to its paper specs there at all so it won't be surprising if these numbers are completely misleading. But if they hold up it doesn't look like Nvidia will need any $900+ card to retake the performance crown cause they haven't lost it yet. I was just commenting on your comparison to GF6 - it just doesnt fit with the current situation (based on DT's numbers).

[Geo]

But that's my point. DT's numbers are based on a $939 BFG watercooled 650mhz core 8800.

Regardless of whether the R600 numbers are dodgy or not.

[trinibwoy]

On 1280x1024 they tested the HD 2900XT getting 47.9fps in Oblivion. However, their latest benchmarks put the HD 2900XT getting 73.4fps in a resoluton of 1920x1200. Same CPU and same game settings.

Yep, it seems all the 2900XT numbers in the XTX preview are different from the original XT preview. I wouldn't read too much into it just yet - different people did the benchmarking on different machines using whatever random methodology the came up with. I'd wait until somebody can come up with an apples to apples comparison of all the cards involved on the same machine.

[Panajev2001a]

Ok Geo

[Geo]

Take the "AMD was stoopid for buying ATI" stuff to one of the many other threads with that subject.

[INKster]

Yep, it seems all the 2900XT numbers in the XTX preview are different from the original XT preview. I wouldn't read too much into it just yet - different people did the benchmarking on different machines using whatever random methodology the came up with. I'd wait until somebody can come up with an apples to apples comparison of all the cards involved on the same machine.

The drivers are different, though:
The benchmarks DailyTech performed yesterday utilized release candidate drivers.
Today's tests used retail drivers ATI released to its board partners.

[leoneazzurro]

The drivers are different, though:

A retail driver performing worse than a beta driver. hmmm :???:

[vertex_shader]

But that's my point. DT's numbers are based on a $939 BFG watercooled 650mhz core 8800.

Regardless of whether the R600 numbers are dodgy or not.

Give us a hint, dodgy or not? :razz:

[Banko]

A retail driver performing worse than a beta driver. hmmm :???:

It's possible that with the beta driver some things weren't rendered correctly, or maybe some things weren't even rendered at all.

[Geo]

Give us a hint, dodgy or not? :razz:

I don't know. If I did I couldn't tell you. :grin:

[IbaneZ]

Ok, I figured it out.

This is just a cruel prank, and DT, AMD and fudo are all involved! And when the 2900 series launches we'll be blown away. There.


It only took two rum and cokes to crack it. :smile:

[leoneazzurro]

It's possible that with the beta driver some things weren't rendered correctly, or maybe some things weren't even rendered at all.

That should have been written in the article, as it is fairly easy to detect such things. I'd suggest AMD to go back to the beta drivers :razz:

[Dooby]

Um, why have they benched at 1280? My monitor can do twice that. I want to see "real" numbers (>1600).

Edit : Dont mind me, just passing through. Hadnt seen the second DT link with the XTX before I made this post...

[Razor1]

Ok, I figured it out.

This is just a cruel prank, and DT, AMD and fudo are all involved! And when the 2900 series launches we'll be blown away. There.


It only took two rum and cokes to crack it. :smile:


And Kyle to that list :smile:

[IbaneZ]

And Kyle to that list :smile:

Ahh yes, I forgot about him.

The R600 is going to be ATI's 5800..

The plot thickens. :cool:

[trinibwoy]

But that's my point. DT's numbers are based on a $939 BFG watercooled 650mhz core 8800.

Hmmm, so you think things would have changed considerably if they had used one of the much cheaper 630/1000 variants? That's 97% of the core clock used by DT.

http://www.newegg.com/Product/Product.aspx?Item=N82E16814143093

That GTX isn't expensive because of its configuration. It's expensive because it comes with an expensive water block + markup for pre-installation etc.

[jimmyjames123]

http://www.dailytech.com/ATI+Radeon+HD+2900+XTX+Doomed+from+the+Start/article7052.htm

XTX OEM vs 8800 GTX.

Some weird results. :???:

These benchmarks are barely stressing these cards at all. At 1280x1024, none of the cards get less than 79 average fps. Even at 1920x1200, none of the cards get less than 52 average fps!

With R600, I would think ATI was more interested in getting higher framerates with very high IQ settings, as opposed to simply super high framerates when the cards are not very stressed in the first place. They had the same line of thought with the R520 and R420 as well.

Let's see the cards really get pushed, and then come to better conclusions. That said, it's unrealistic to expect the 512MB XT model in particular to outperform a 768MB 8800 GTX running at 650Mhz core and 2.0Ghz mem. The closer price point competition to the XT would be the 640 MB 8800 GTS. Interesting that today I am seeing pricegrabber links where the 640MB 8800 GTS is selling for $399.

[overclocked_enthusiasm]

It appears now that the simple conclusion as to why ATI didn't release the 2900 XT earlier is because they couldn't beat the 8800 GTX. That means that their cover story about launching the whole R600 stack at once was misleading at best and complete BS at worst. With the delays of the other RV6xx and R6xx cards it seems that this PR cover story has fallen apart with the most glaring failure of all being the likely canning of the 2900 XTX because it just sucks.

The theory (which I hyped and supported) of a unified launch showcasing the Barcelona, R600 and RD790 also seems to be not the case since the 2900 XT can't beat the 8800 GTX. Therefore the "WOW" factor of showcasing Barcelona is lost because the GPU is second best. So I am going to officially say I was wrong on that but it was the ONLY way I could get the PR mixed messages from AMD to sort out.

In truth they didn't release R600 months ago because they either didn't want to be second best, they thought they had time to beat 8800 GTZ with 2900 XTX or they simply couldn't becuase of driver or manufacturing problems. The VERY sad part is that none of these reasons jives with what was coming out of the ATI or AMD camp for the past many months and I consider their PR and management rhetoric to have been, in hindsight, disingenuous at best.

AMD/ATI has lost ALL credibility with me at this point. Management has given repeated assurances on product launch windows, product launch strategies, market share, gross margins, ASPs and other financial/product releated issues that have ALL been recently debunked. You can put me in "SHOW ME" camp from now on concerning ANYTHING to do with either AMD or ATI. Again I ask the question...why does Orton still have a job?

[Lezmaka]

We all know Wattage is just part of the answer.
Nvidia recommends 26amps for a GTS, 30amps for a GTX.

How much for this R600 XT in that respect ?

Wattage is amps x voltage. They recommend 26/30 amps because of all the crap power supplies out there. 26a x 12V = 312W and 30a x 12V = 360W. You should be fine if you have a good quality power supply and not just one that says ooh! look! 550W!!!!

This system had a peak DC power consumption (http://www.silentpcreview.com/article28-page4.html) of 256W while running PCMark05. That's not just on 12V, but everything (3.3V, 5V, 12V).

Intel Pentium D950, overclocked / overvolted 10% (Presler, 3.74GHz)
Asus P5LD2-VM motherboard
4 x 1024MB Corsair DDR2-6400 RAM
ATI Radeon X1950XTX-512 PCIe graphics
Hitachi Deskstar 7K80 80GB hard drive
WD Raptor WD1500ADFD 150GB 10krpm HDD
Seasonic SS-350ET ATX PSU
1 x 120mm fan

Even if the 2900XT uses all 225W it can get, that would still leave 183W on 12V rails for CPU, HDD, etc. Anyway, page 3 of the article I linked to has good info on "how much is enough." Read that and make your own decision if your PSU will be good enough.

[Sound_Card]

-Fudo claims that AMD stated that the XTX will be slower than the GTX. Dispite him not being invited to tusinia and not intending.

-Daily techs performance figures place the x2900xtx at a slower speed than the x1950xtx.

-Daily tech bench marks the x2900xt pre maturaly using sub par setting's and mid range card buyer's resolution making a claim that they did not have a better monitor. Like that is to be believed.:roll:

-fundamental flaws are seen in dailytechs performance figures.

-AMD gives R600's to people that don't sign NDA's? :shock:

-Since dailytech does not sign NDA's, why is their not a full spec layout of the chip? They quickly give numbers to show, but they don't have anything to back their claim.

-review was done in less than a day. Kind of a joke to me.:roll:




Anyone remember the truck loads of bad news just weeks before G80 came out?

[Geo]

Hmmm, so you think things would have changed considerably if they had used one of the much cheaper 630/1000 variants? That's 97% of the core clock used by DT.

http://www.newegg.com/Product/Product.aspx?Item=N82E16814143093

That GTX isn't expensive because of its configuration. It's expensive because it comes with an expensive water block + markup for pre-installation etc.

They would appear to offer it for a reason. And unless someone tells us that the presumed Ultra is now 65nm, one wonders how fast or how massively (and expensively) cooled it will have to be in order to be clocked any higher than that.

[satein]

:smile: -Fudo claims that AMD stated that the XTX will be slower than the GTX. Dispite him not being invited to tusinia and not intending.

-Daily techs performance figures place the x2900xtx at a slower speed than the x1950xtx.

-Daily tech bench marks the x2900xt pre maturaly using sub par setting's and mid range card buyer's resolution making a claim that they did not have a better monitor. Like that is to be believed.:roll:

-fundamental flaws are seen in dailytechs performance figures.

-AMD gives R600's to people that don't sign NDA's? :shock:

-Since dailytech does not sign NDA's, why is their not a full spec layout of the chip? They quickly give numbers to show, but they don't have anything to back their claim.

-review was done in less than a day. Kind of a joke to me.:roll:




Anyone remember the truck loads of bad news just weeks before G80 came out?


Probably, it is like giving a monkey a diamond. There were too many bad news around at the moment which sounds too strange to believe.

Hope we would get a mature review out soon, and I would love to read on how B3D investigate this doggy-delay-R600

[trinibwoy]

They would appear to offer it for a reason. And unless someone tells us that the presumed Ultra is now 65nm, one wonders how fast or how massively (and expensively) cooled it will have to be in order to be clocked any higher than that.

I'm still not following you. Why is the Ultra relevant at all if AMD can't beat the GTX?

[INKster]

I'm still not following you. Why is the Ultra relevant at all if AMD can't beat the GTX?

Keeps the prices jacked up in the high-end.

[SugarCoat]

Anyone remember the truck loads of bad news just weeks before G80 came out?

Nope. I remember the specs were leaked in the weeks just before it came out and looked so out of wack from what anyone expected people had a hard time believing them. Course they turned out mostly true.

[Unknown Soldier]

People tend to forget the G80 has Vista driver problems. Can anyone say with 100% certainty that the G80(and future Ultra) and family now have a flawless Vista(32and 64bit) driver?

As for the R600 drivers, it's still needs to be called into question.

US

[Natoma]

-Daily techs performance figures place the x2900xtx at a slower speed than the x1950xtx.

That alone leads me to believe these #s are BS. Even the NV30 trounced the GF4Ti while being trounced by the 9700 Pro. And we're to believe the R600 is so bad it's no faster than a x1950xtx?

Please.... :roll:

[INKster]

-Fudo claims that AMD stated that the XTX will be slower than the GTX. Dispite him not being invited to tusinia and not intending.

-Daily techs performance figures place the x2900xtx at a slower speed than the x1950xtx.

-Daily tech bench marks the x2900xt pre maturaly using sub par setting's and mid range card buyer's resolution making a claim that they did not have a better monitor. Like that is to be believed.:roll:

-fundamental flaws are seen in dailytechs performance figures.

-AMD gives R600's to people that don't sign NDA's? :shock:

-Since dailytech does not sign NDA's, why is their not a full spec layout of the chip? They quickly give numbers to show, but they don't have anything to back their claim.

-review was done in less than a day. Kind of a joke to me.:roll:



You are so wrong about that...

Anh Huynh (DailyTech Staff)
Before anyone asks, the quality settings for the games were as follows:

Company of Heroes - High shader quality, High model quality, Anti-aliasing enabled (in game), Ultra texture quality, high quality shadows, high quality reflections, Post processing On, High building detail, High physics, high tree quality, High terrain detail, Ultra effects fidelity, Ultra effects density, Object scarring enabled and the model detail slider all the way to the right.

F.E.A.R. - 4x FSAA (in game), maximum light details, shadows enabled, maximum shadow details, soft shadows enabled, 16x anisotropic filtering, maximum texture resolution, maximum videos, maximum shader quality.

Half Life 2: Episode 1 - High model detail, high texture detail, high shader detail, reflect all water details, high shadow detail, 4x multi-sample AA (in-game), 16x anisotropic filtering, v-sync disabled, full high-dynamic range.

Elder Scrolls IV Oblivion – Preset ultra quality settings. The scene used was where you’re at the end of the tunnel and run outside for a little bit.

Does this look like "sub par settings" to you ?

Also, AMD doesn't give cards because they don't manufacture them, partners do.

[Geeforcer]

-Daily techs performance figures place the x2900xtx at a slower speed than the x1950xtx.

You can't just compare numbers from two different reviews. Were the settings exactly the same? Timedemos the same? Etc?

-Daily tech bench marks the x2900xt pre maturaly using sub par setting's and mid range card buyer's resolution making a claim that they did not have a better monitor. Like that is to be believed.:roll:

Why, how many monitors do people usually lug around?

Anyway, as far as DailyTech is concerned.. their pre-launch 8800 numbers (http://www.dailytech.com/article.aspx?newsid=4812)were pretty much spot on.

Anyone remember the truck loads of bad news just weeks before G80 came out?

Other then faulty resistors, no, not really.

[Sound_Card]

Nope. I remember the specs were leaked in the weeks just before it came out and looked so out of wack from what anyone expected people had a hard time believing them. Course they turned out mostly true.


I'm talking about the reports of G80 not being unifide, DX10 stamped on 7900GTX, perofrmance slightly better than the x1900xt, horrible IQ.


All this was about a month before release. What I see is a bunch of pre mature gossip web sites making qustionable claims.

[trinibwoy]

Hope we would get a mature review out soon, and I would love to read on how B3D investigate this doggy-delay-R600

I'm just glad R600 is out now so these B3D folks can stop using its delay as an excuse for the postponement of a DX10 investigation :roll: :razz:

[Geo]

I'm still not following you. Why is the Ultra relevant at all if AMD can't beat the GTX?


I have no idea. You were the one who asserted its relevance to the thread upstream by saying AMD would have no answer for NV's top two cards. Or did you have something else in mind?

[trinibwoy]

Keeps the prices jacked up in the high-end.

Thank you INKster for further confusing my conversation with geo :) I'm not talking about relevance from Nvidia's perspective.

Geo seems to think think that the Ultra is in some way key to Nvidia claiming the performance crown. I'm asking why he thinks the Ultra is relevant based on DT's numbers. There are GTX's out there at those speeds.

[Mize]

The problem with the "it's BS" theory is where's the incentive for DT? A few days of high hit counts followed by being a laughing stock?

Assuming these numbers hold up the, if AMD are lucky it's they're 5800. If they're not it's their Voodoo 5-5500. :(

[trinibwoy]

I have no idea. You were the one who asserted its relevance to the thread upstream by saying AMD would have no answer for NV's top two cards. Or did you have something else in mind?

Exactly, I was trying to demonstrate that your GF6 comparison was poor because that was flagship vs flagship. Today we're looking at flagship vs two-cards-down if these numbers hold.

[Sound_Card]

You can't just compare numbers from two different reviews. Were the settings exactly the same? Timedemos the same? Etc?



Oh come on!

The setting are listed in the link that I kindly provided. It is also here in this very thread.

2900XTX paired with a Intel kentsfield is getting the same frames as the x1950xtx paried with a Intel conroe.

More data is needed, and the smoke needs to be clear.

[Geo]

Exactly, I was trying to demonstrate that your GF6 comparison was poor because that was flagship vs flagship. Today we're looking at flagship vs two-cards-down if these numbers hold.

So you've changed your mind again and now Ultra is a relevant part of the discussion? I'm getting whiplash here. . . .

[HKS]

[QUOTE=Sound_Card;976401]
-AMD gives R600's to people that don't sign NDA's? :shock:

-review was done in less than a day. Kind of a joke to me.:roll:
[QUOTE]
Somehow DailyTech managed to get a 8800GTX a week before launch and run benchmarks with same drivers as the official benchmarks a week later.

The numbers in their 8800GTX "preview" was pretty accurate...

The way I understand it is that they are testing on the cards AnandTech recive, that's the reason the short time to do benchmarks.

[Natoma]

The problem with the "it's BS" theory is where's the incentive for DT? A few days of high hit counts followed by being a laughing stock?

Assuming these numbers hold up the, if AMD are lucky it's they're 5800. If they're not it's their Voodoo 5-5500. :(

When has the high end next-gen solution of any hardware company not outperformed it's previous generation offering by a significant margin at the same settings? That's what has my red flags raised right now. That wouldn't be NV30 or even Voodoo 5-5500. This would be Voodoo 3 vs Voodoo 2 SLI all over again, and we know what happened there.

I'm having a difficult time believing these #s because if they are indeed true, this would be a blunder of enormous proportions. Far worse than anything a major graphics company has done in the past 5-7 years.

[XMAN26]

http://www.dailytech.com/article.aspx?newsid=7052

Things dont look good if these numbers are true.

[Natoma]

http://www.dailytech.com/article.aspx?newsid=7052

Things dont look good if these numbers are true.

Many moons ago. http://forum.beyond3d.com/showthread.php?p=976084#post976084

[INKster]

http://www.dailytech.com/article.aspx?newsid=7052

Things dont look good if these numbers are true.

That was almost 150 posts ago... today.

http://forum.beyond3d.com/showpost.php?p=976084&postcount=2801

:wink:

[jimmyjames123]

http://www.dailytech.com/article.aspx?newsid=7052

Things dont look good if these numbers are true.

Umm hello, that's what we've been discussing for the last three or four pages of this thread :D

[Evildeus]

Umm hello, that's what we've been discussing for the last three or four pages of this thread :D
5 in fact :lol:

[Natoma]

I beat all you bitches in pointing out XMAN's late-to-the-party snafu. :wink:

I'm first! :D

[SugarCoat]

I'm talking about the reports of G80 not being unifide, DX10 stamped on 7900GTX, perofrmance slightly better than the x1900xt, horrible IQ.


All this was about a month before release. What I see is a bunch of pre mature gossip web sites making qustionable claims.

no actually im pretty sure that was months before release, especially the DX10 bolted on statement which was quite awhile before the release. The mostly accurate leaked specs was in the time frame you mentioned, "weeks before release".

[vertex_shader]

Somehow DailyTech managed to get a 8800GTX a week before launch and run benchmarks with same drivers as the official benchmarks a week later.

The numbers in their 8800GTX "preview" was pretty accurate...

The way I understand it is that they are testing on the cards AnandTech recive, that's the reason the short time to do benchmarks.

Don't forget dailytech/anandtech is nv pro sites, anandtech was the first come out with the performance/watt ratio after the r520 released :smile:

I not say its can't be true what dailytech write, but very unprofessional the way how they tested the cards, so i take the whole test as a grain of salt.

I wait for may 14th than i have a clear picture for any conclusion :smile:

[HKS]

Don't forget dailytech/anandtech is nv pro sites, anandtech was the first come out with the performance/watt ratio after the r520 released :smile:

I realy hope this was ironic...
AnandTech usualy have very good tests, and they were not kind when testing 8600/8500

[Unknown Soldier]

I beat all you bitches in pointing out XMAN's late-to-the-party snafu. :wink:

I'm first! :D

whahahah .. liar liar pants on fire :D

http://www.bilgiustam.com/?p=125

HD2900XT faster than the 8800GTX in Crysis?

What nice is their specs are almost the same as mine :)

Test Sistemi:
Intel Core 2 Duo E6600 (2.4Ghz) (Mine: QX6600)
Asus P5W DH
2×1GB PC2-6400 (5-5-5-12)

US

[IbaneZ]

Don't forget dailytech/anandtech is nv pro sites

I believe those days are over. Are there any pro sites left? :wink:

[Geeforcer]

I'm talking about the reports of G80 not being unifide, DX10 stamped on 7900GTX, perofrmance slightly better than the x1900xt, horrible IQ.


All this was about a month before release. What I see is a bunch of pre mature gossip web sites making qustionable claims.

That wasn't just weeks before G80 came out and it wasn't on DailyTech - I'd know since I complied the collection of all INQ stories that were off target. Here are the Dailytech stories:

G80 lenght and power. (http://dailytech.com/article.aspx?newsid=4442)

Specks (http://www.dailytech.com/article.aspx?newsid=4441)

More specks (http://www.dailytech.com/article.aspx?newsid=4450)

[Geo]

whahahah .. liar liar pants on fire :D

http://www.bilgiustam.com/?p=125

HD2900XT faster than the 8800GTX in Crysis?

What nice is there specs are almost the same as mine :)

Test Sistemi:
Intel Core 2 Duo E6600 (2.4Ghz) (Mine: QX6600)
Asus P5W DH
2×1GB PC2-6400 (5-5-5-12)

US

Oy! On the one hand, very interesting. On the other, possibly even less deserving of reliance than DT's scores, as at least DT's scores will be easily confirmable/refutable by reviewers. Unless Crytek has released a demo I'm not aware of, one can't say that for these alleged Crysis benchies. . .

[HAL]

http://www.vr-zone.com/?i=4931
:wink:

[jimmyjames123]

whahahah .. liar liar pants on fire :D

http://www.bilgiustam.com/?p=125

HD2900XT faster than the 8800GTX in Crysis?

What nice is their specs are almost the same as mine :)

Test Sistemi:
Intel Core 2 Duo E6600 (2.4Ghz) (Mine: QX6600)
Asus P5W DH
2×1GB PC2-6400 (5-5-5-12)

US

Although this could very well be the case, note that they are using older Forceware drivers on the 8800 GTX. The new ones could make a difference. Also, note that Dailytech was testing a 8800 GTX at 650Mhz core and 2.0 Ghz mem, which is quite a bit faster than stock GTX.

[INKster]

http://www.vr-zone.com/?i=4931
:wink:

So, the HD2900 XT core clock is even slower ? (5 MHz less).

[IbaneZ]

http://www.vr-zone.com/?i=4931
:wink:

I love you! :lol:

That's on wednesday then. OMG!

[Geo]

http://www.vr-zone.com/?i=4931
:wink:

:shock: That would be roughly akin to pushing up an invasion by two weeks.

[Mize]

:shock: That would be roughly akin to pushing up an invasion by two weeks.

Or a surrender ;)

[INKster]

:shock: That would be roughly akin to pushing up an invasion by two weeks.

An invasion with a platoon, instead of an army, perhaps ? (in keeping with the military analogies :D)

[jimmyjames123]

This doesn't surprise me. I see no reason why the HD 2900 XT 512MB version will not compare favorably to the 8800 GTS 640MB version. So if both are slated for the same price point right now (noting that the 8800 GTS 640MB version even now has a street price of $399), then why not have a preview that is a couple weeks earlier than planned?

[vertex_shader]

I realy hope this was ironic...
AnandTech usualy have very good tests, and they were not kind when testing 8600/8500

Nope, im serious.
8500/8600 is another story, nv want sell more 8800gts 320mb, so to conclusion 8500/8600 is a garbage in this price, helping nv.

Most of the sites are manipulativ, ati or nv way, nv have more manipulativ website in there pocket.
May 14th we see eveything, than we can say clearly r600 is a blow or not (not so clearly because dx10 applications still far away).

[Unknown Soldier]

:shock: That would be roughly akin to pushing up an invasion by two weeks.

Now it begs to ask the question, has B3D already received their HD2900XT/XTX?? :D

US

[INKster]

XT on the 2nd, Ultra on the 1st:
http://www.vr-zone.com/?i=4930

[vertex_shader]

http://www.vr-zone.com/?i=4931
:wink:

So this this AMD answer for the FUD campaign what is started against amd and r600, or this is another BS only :smile:

Next FUD will come in may 2nd, "AMD paper launched the r600 series" :wink:

[jimmyjames123]

I don't completely understand why people are saying that R600 in general is a complete failure. Maybe the XTX is a failure if it doesn't outperform the GTX. However, the 512MB XT model seems to perform as good if not better than the 640MB GTS, for a similar price point.

Surely the 512MB XT model is a success for consumers if it performs as good if not better than the 640MB GTS, and sells for a similar price? NVIDIA partners would have to be willing to sell the 768MB 8800 GTX for about $449 or lower for it to really trump the 512MB XT model.

[Geeforcer]

Oh come on!

The setting are listed in the link that I kindly provided. It is also here in this very thread.

2900XTX paired with a Intel kentsfield is getting the same frames as the x1950xtx paried with a Intel conroe.

More data is needed, and the smoke needs to be clear.

Let's see: The benchmarks you provided for comparison were CoH from TweakTown and FEAR from Tom's.

DTs CoH settings: High shader quality, High model quality, Anti-aliasing enabled (in game), Ultra texture quality, high quality shadows, high quality reflections, Post processing On, High building detail, High physics, high tree quality, High terrain detail, Ultra effects fidelity, Ultra effects density, Object scarring enabled and the model detail slider all the way to the right.

TweakTown CoH settings: ???

See a problem here?

[overclocked_enthusiasm]

XT on the 2nd, Ultra on the 1st:
http://www.vr-zone.com/?i=4930

AMD CEO Hector Ruiz said in the CC last week "we don't do soft launches".

[Geeforcer]

:shock: That would be roughly akin to pushing up an invasion by two weeks.

Seriously, AMD's launch strategy at this point gets five OMGWTFBBQs from me.

[fallguy]

whahahah .. liar liar pants on fire :D

http://www.bilgiustam.com/?p=125

HD2900XT faster than the 8800GTX in Crysis?

What nice is their specs are almost the same as mine :)

Test Sistemi:
Intel Core 2 Duo E6600 (2.4Ghz) (Mine: QX6600)
Asus P5W DH
2×1GB PC2-6400 (5-5-5-12)

US

Why is it that nobody else has this demo?

[overclocked_enthusiasm]

Seriously, AMD's launch strategy at this point gets five OMGWTFBBQs from me.

Well, with the 2900 XTX canned and the other RV6xx and R6xx parts delayed, wht's the point in waiting another 2 weeks especially if the NDA's are starting to crack? It is a comedy of errors at this point and the sooner the card(s) are released the sooner they can start hyping how great the R650 is going to be.

[satein]

AMD CEO Hector Ruiz said in the CC last week "we don't do soft launches".

Probably, it is because of so many leaks and it seems uncontrollable of 3-weeks silent NDA.

Putting it ahead would not be a serious damage control for AMD/ATi at this point through...

[3dilettante]

I figured I'd muse a little on the idea that R600 is intended for use on other kinds of workloads, and whether this will even be usable in a game situation.

Would it go for a coarse kind of task switching, where at any given time R600 is only working on graphics, or only working on physics, etc?

Could it go for something with finer granularity, as in one partition of the chip can be running audio or video code while the rest is working on graphics.

The coarse method seems simpler, though possibly more wasteful with so much of the core idling.

The finer method may allow for better utilization up-front, but if ROPs and TMUs are clustered, doing this can cause R600 to lose large fractions of its pixel and sampling fill rate.
If a physics thread barely touches any of that hardware, it is lost to graphics threads that could use it.

[INKster]

AMD CEO Hector Ruiz said in the CC last week "we don't do soft launches".

Honestly, i'd rather believe VR-Zone's story about the May 2nd anticipation + "soft launch" right now.
Hector is no Jerry Sanders...

[Unknown Soldier]

Why is it that nobody else has this demo?

ask me i'm god.

US

[nutball]

So if both are slated for the same price point right now (noting that the 8800 GTS 640MB version even now has a street price of $399), then why not have a preview that is a couple weeks earlier than planned?

Because... it looks like AMD marketing doesn't know what the **** it's doing? Running around like headless chickens? This "launch" has had more ins and outs and ups and downs than a cheap porn movie.

[Unknown Soldier]

Because... it looks like AMD marketing doesn't know what the **** it's doing? Running around like headless chickens? This "launch" has had more ins and outs and ups and downs than a cheap porn movie.

Frankly, I beleive AMD dropped the ball when they delayed in March already.

US

[nutball]

Frankly, I beleive AMD dropped the ball when they delayed in March already.


Yes, that's what I was meaning really. The whole episode, not just this latest twist. Last minute tweaks happen, but the last six months just smack of vacant stares around conference tables.

[trinibwoy]

So you've changed your mind again and now Ultra is a relevant part of the discussion? I'm getting whiplash here. . . .

Not sure what I've said that could possibly give you that impression. I think you're focusing two much on the fact that I said that the GTS was two products down - the GTX and Ultra being its superiors. My main point is that from all appearances the XT (AMD's flagship) won't be going up against Nvidia's current flagship GTX, hence the Ultra is irrelevant. Was just trying to understand how that was similiar to the GF6 situation in your mind.

But this is all based on numbers that are highly suspect so I'll stop nagging you now.

[AlexV]

Coz nobody else thought it would be cool to pull something out of their dark dark anal cavity?Probably not, since it shows ATi as OMGWTFTEHSAUCE and thus must be true. Personally, from a business standpoint, I do feel that the x2900xt will do wonderful things for consumers...dunno if for AMD it`ll have a similar effect, but for us it should usher in some very very nice price adjustments.

[Kombatant]

Why is it that nobody else has this demo?That guy probably has a Alan Wake demo as well, let's raid his offices :cool:

On a serious note, I still see some stuff out there that have no merit. Patience.

[trinibwoy]

I don't completely understand why people are saying that R600 in general is a complete failure. Maybe the XTX is a failure if it doesn't outperform the GTX. However, the 512MB XT model seems to perform as good if not better than the 640MB GTS, for a similar price point.

Well you can't avoid the architecture to architecture comparisons. You seem to be implying that they can just set a low price on their fastest card and say "Hey look, we're fastest at $xxx". I think the price point of the XT is a lot more telling than any numbers from DT. If they thought it could compete favorably at a higher price point this move would boil down to initiating a price war. Exactly why would they do that?

[Natoma]

That guy probably has a Alan Wake demo as well, let's raid his offices :cool:

On a serious note, I still see some stuff out there that have no merit. Patience.

Talking about the DT benchies I hope. Otherwise XTX is a damn dog.....

[Slappi]

Talking about the DT benchies I hope. Otherwise XTX is a damn dog.....

Well he just posted at Rage about DTs cred so maybe that is what he meant.

[SugarCoat]

I don't completely understand why people are saying that R600 in general is a complete failure. Maybe the XTX is a failure if it doesn't outperform the GTX. However, the 512MB XT model seems to perform as good if not better than the 640MB GTS, for a similar price point.

Surely the 512MB XT model is a success for consumers if it performs as good if not better than the 640MB GTS, and sells for a similar price? NVIDIA partners would have to be willing to sell the 768MB 8800 GTX for about $449 or lower for it to really trump the 512MB XT model.

you dont see the glaring problem in your view? The cards launching almost 6 months after nVidia and if it offers nothing really earth shattering thats really about as bad as you can get short of performing worse in many titles as well.

The XT is also launching at a reactive price DUE to its late introduction so i think you're being far too lenient. IF the card had launched within a reasonable time frame to the competition it would easily have been a $499 part rather then a $399. They've basically been forced into the $399 pricepoint, that you think looks so good, due to how late the launch is.

We're also talking about the flagship card as well until an XTX shows its face in quantity, the same goes for any 8800Ultra, the GTX is still the flagship card. This launch, at least from this distant perspective that leaked information is offering, is turning out to be as exciting as the X800XT was against the 6800Ultra (not a huge performance gap or large advantage in IQ), except this time instead of launching withing a few weeks of eachother ATi is coming in 6 months later which makes it look terrible.

[Jawed]

I figured I'd muse a little on the idea that R600 is intended for use on other kinds of workloads, and whether this will even be usable in a game situation.

Would it go for a coarse kind of task switching, where at any given time R600 is only working on graphics, or only working on physics, etc?

Could it go for something with finer granularity, as in one partition of the chip can be running audio or video code while the rest is working on graphics.

The coarse method seems simpler, though possibly more wasteful with so much of the core idling.

The finer method may allow for better utilization up-front, but if ROPs and TMUs are clustered, doing this can cause R600 to lose large fractions of its pixel and sampling fill rate.
If a physics thread barely touches any of that hardware, it is lost to graphics threads that could use it.
Virtual memory fragment aware cache (http://v3.espacenet.com/textdoc?DB=EPODOC&IDX=US2007073996&F=0)

[0051] Context switching includes changing out one page table for another. This entails saving state for non-current contexts. Context switching avoids inefficient down-time when one page table is being modified. Clients that run together are grouped to make the context switching more efficient. In various embodiments, a context field (not shown) allows contexts to be specified on a per-memory-access-request basis.

[0112] Page table units 818 support up to eight active physical page table contexts simultaneously. A first page table context 0 is reserved for the so-called system context. The remainder can be assigned at will, but in one embodiment various clients each have their own independently managed page table contexts. Types of clients include for example, video, multi-media, and 3D graphics clients. In one embodiment, multimedia clients use a context that is independent from the 3D graphics client context.

Jawed

[ban25]

Surely the 512MB XT model is a success for consumers if it performs as good if not better than the 640MB GTS, and sells for a similar price? NVIDIA partners would have to be willing to sell the 768MB 8800 GTX for about $449 or lower for it to really trump the 512MB XT model.

A success for consumers how? Consumers have had 'as good' a card since last year...

[neliz]

Not sure if someone posted it yet. Just got an update on my pre-order, it got postponed 'till May 28th

[Geeforcer]

neliz, who did you pre-order from?

[jimmyjames123]

Not sure what I've said that could possibly give you that impression. I think you're focusing two much on the fact that I said that the GTS was two products down - the GTX and Ultra being its superiors. My main point is that from all appearances the XT (AMD's flagship) won't be going up against Nvidia's current flagship GTX, hence the Ultra is irrelevant. Was just trying to understand how that was similiar to the GF6 situation in your mind.

But this is all based on numbers that are highly suspect so I'll stop nagging you now.

Even if there is no XTX, the Ultra is relevant in the sense that it steals some attention from AMD's new launch.

[Cuthalu]

Has DT mentioned anything about power consumption or power connectors? If it indeed does consume a lot more power than GTX while being slower at the same time, then it'll suck big time.

[INKster]

Has DT mentioned anything about power consumption or power connectors? If it indeed does consume a lot more power than GTX while being slower at the same time, then it'll suck big time.

Well, at least the air sucking part of it is similar (if the fan is indeed the same). :D

As for the power requirements, yes, that's true.
I'm still waiting to learn about the recommended amperage of the XT.

[jamis]

Personally, from a business standpoint, I do feel that the x2900xt will do wonderful things for consumers...dunno if for AMD it`ll have a similar effect, but for us it should usher in some very very nice price adjustments.
And just think about those 65nm refresh parts! Now if the rumours are true and both the NV and AMD refreshes are just shrinks and NV gets it's refresh out around the same time as AMD, they will have to compete with prices again. So we'll be getting 65nm R650 with high clocks , low power consumption, low noise and even lower prices!:grin:
Provided of course that AMD is still in business..:lol:

[ERK]

Until we know the architecture for all kinds of buffer fetches (vertex, texture, constant) it's hard to judge how bandwidth-demanding the generic concept of "texturing" is. As I said earlier, D3D10's approach prolly means that you can't judge capability from the number of TMUs.

As much as anything, R600 might be more focussed on D3D10. Wasn't R300 seen as nothing more than a "slightly faster DX8", with AF, AA and SM2 taking centre stage? How long did ATI have to wait for that to pay-off? :lol:

Having said all this, 16 TMUs at anything like the same "rate" as R5xx seems rather stingy. 120GB/s+ for DX9 apps really is disproportionate if the TMUs aren't up to it - and CrossFire R5xx shows that there's plenty more performance to come in apparently texturing-limited games. Let alone G8x, which one way or another, seems to show the same.

16 TMUs, considered on their own, would make R600 a 4:1 ALU:TEX architecture. I've long thought that R5xx would prolly have performed the same if it was only 2:1 instead of 3:1. For DX9 apps, 4:1 may well just be beyond the pale.

But under D3D10, with the variety of buffers seen by the programmer, the "ALU:TEX" ratio may fall, it might be 2:1, if there are pipes dedicated to point sampling or constant buffer fetches. Xenos has 16 dedicated point samplers (which are mostly used for vertex buffer fetches) but they can also be used to perform non-filtered "texture fetches" or constant fetches (guessing about the latter).

Might be better to think of ALU:Fetch ratio.

Jawed

Not asking you, specifically, Jawed, but quoting for context, and assuming an architecture somewhat based on Xenos as above...

How on God's green earth can they take a 232M transistor chip, add 16 more shader ALUs (vec 4+1 or vec5) to the existing 48, and end up with 700M+ trannies????

For cripes' sake! Even adding a GS and more int functionality...surely D3D10 compliance does not require 400M transistors. :???:

I'm really missing something here.
(I realize the 232M figure does not include the ROPs, etc. but still...)

ERK

[jimmyjames123]

Well you can't avoid the architecture to architecture comparisons. You seem to be implying that they can just set a low price on their fastest card and say "Hey look, we're fastest at $xxx". I think the price point of the XT is a lot more telling than any numbers from DT. If they thought it could compete favorably at a higher price point this move would boil down to initiating a price war. Exactly why would they do that?

One design decision that is in AMD's favor price-wise for the XT is the fact that they can get away with placing the 512MB XT against the 640MB 8800 GTS, since performance of the XT is supposed to be as good or better in comparison. Having 128MB less RAM is some ground for cost savings, don't you think? Given the apparent failure of the XTX to outpace the GTX at the very high end, then the next best thing is to have a price-to-performance advantage at lower price point(s). It's not like AMD partners are making no money selling the XT at $399-$449. The fact of the matter is that having less RAM on board in comparison to a similarly performing card from the competition gives them some cost savings in that area, so they can price it very competitively while still making money on it.

Obviously if the XT could compete against the 8800 GTX, they would be pricing it much higher. Maybe it can't quite reach that level of performance, so it has to be priced lower than the GTX, which is no big deal given the fact that it has 256MB RAM less in comparison.