The NEXT LAST R600 Rumours & Speculation Thread [Archive] - Page 2

rwolf

03-Mar-2007, 02:11

Where is the fundimental change of how graphics is programmed?

Cap elimination is great it will actually reduce the amount of work for engineers, no extra features for dx outside the spec.

Virtual memory, does this really have much to do with over all speed of the graphics card, think developers will really want to use virtual memory when making thier games

API object overhead reduction, I don't see where that really fits in its accessible to any dx10 graphics card

HLSL10, and SM 4.0 and geometry shaders, this is the only area that might be of any concern which I mentioned anything that has been done so far by nV, in the g80 which has increased Dx9 performance will also show up in Dx10 games. Now if we had something to compare to would be nice :)

standard storage formats this is accessible to all dx10 cards as well.

- geometry shader and the ability to create ploygons.
- Substantially reduced API object overhead

These two feature alone will solve the problem with applications being completely CPU bound.

- no cap bits all features in hardware eliminating multiple code paths in games.
- virtual memory
- Unified instruction sets (HLSL 10)
- Shader model 4.0
- Standard Storage Formats

These will make games easier to code and reduce development time.

Razor1

03-Mar-2007, 02:29

hmm well all dx10 gpu's have those features, and that wasn't what we were talking about, which features of dx10 could be f*cked up or done better is what I'm talking about, if a GPU is a good dx9 performer, all of these features in dx10 will also run well, so only the additional features like GS is really the only area I can think of that would make a difference. We already see that the load balancing on the g80 for the unified shaders is doing very well but this is possibly another area AMD could have improved since its they second generation USA design. Another area possibly branching, the g80 has very good branching performance, this should be carried over to dx10, but its a comparative situation, the r600 might have better.

Razor1

03-Mar-2007, 02:34

Who ever wrote that article is completely clueless.

Sapphire will never make Geforce cards.
The parent company (PC Partner) has an equivalent brand setup just for selling Nvidia cards called "Zotac".
I believe there are a few of those cards around.

They aren't talking about PC Partner and daughter or sister companies, probably used the wrong term, I'm pretty sure they mean AIB partners.

Ailuros

03-Mar-2007, 06:36

If it doesn't win more than it loses at 8xaa/16xaf at launch and the next few months past that I think it'd be fair to call it a disappointment. How much of a disappointment would depend on just how badly it fails to meet that standard.

If best IQ doesn't improve vs X1k, I'd be disappointed. . .but this is unlikely to happen, as they've nearly certainly at a minimum bumped up from 6x to 8x AA. I still have hopes they've got something else in their pocket on the IQ front, but no evidence to support it.

If they're not more competitive in the midrange at launch than the previous two generations I'd be disappointed as well. Three generations in a row of having your hat handed to you in the fat part of the market is not a good thing in the least, "halo" or not. I'm pointing at the $159-$249 range here.

A quick reality check will tell you that any GPU will have a hard time in upcoming games to use even 8x MSAA in a decent resolution.

turtle

03-Mar-2007, 06:54

A quick reality check will tell you that any GPU will have a hard time in upcoming games to use even 8x MSAA in a decent resolution.

Why? It looks like 4x will be usable on G80 for the time-being in DX9 (and perhaps DX10), why wouldn't the R600 with the possibility of literally twice the bandwidth (179+GBPS vs 86.4) if using 2800mhz GDDR4 be an exception for 8x?

I mean granted, new games are taxing, and that might be pushing it to the brink of what to expect, but it is at least possible...especially considering ATi has done a more efficient job with AA/AF in the past using the same amount of BW as nvidia, imagine them having 2x as much. :twisted:

Sound_Card

03-Mar-2007, 07:15

Could we see RD790 (http://www.chilehardware.com/articulo_1707&pageid=2.html)with R600?

:smile:

DemoCoder

03-Mar-2007, 08:56

- geometry shader and the ability to create ploygons.

Likely not to be used very much in the next 2 years I predict.

- Substantially reduced API object overhead

Irrelevent to underlying HW, so not a factor in G80 DX10 performance vs R600 DX10 performance.

- no cap bits all features in hardware eliminating multiple code paths in games.
- virtual memory
- Unified instruction sets (HLSL 10)
- Shader model 4.0
- Standard Storage Formats

These will make games easier to code and reduce development time.

Yes, but all of this is irrelevent to the original point which is that DX9 performance is likely to predict DX10 performance for the most part. There is a sort of FUD myth going around that somehow the G80 isn't a true DX10 card and somehow DX10 performance will such much like the NV30 was great at DX8 but sucked on real DX9 workloads. In other words, when benchmarks show the G80 and R600 doing about the same on DX9 workloads within margins (maybe +/- 15%), the rallying cry of f*nb*ys will be "but this is not really a fair comparison, wait until real DX10 games come out which will really show the difference between the GPUs!" To which I say, it won't make a big difference.

There are really only 2 DX10 features I can think of that don't have equivalents in DX9 that can be implemented "badly" and hence kill performance if a game used them heavily: Geometry Shaders and Stream Out.

But looking at the CUDA architecture, it is reasonable to expect that both of these will perform as expected on the G80, although I think first-generation Geometry Shaders will essentially remain a toy. There are limitations in DX10 on GS that reduce the usefulness compared to developers using middleware to code to CUDA/CTM directly.

IMHO, I think alot of geometry amplification techniques dovetail with GPGPU techniques and hence, you'll see developers going the Brook/GPGPU route. For example, to do the kind of procedural geometry synthesis you see in something like Speed Tree, you're not going to do it in GS.

Twinkie

03-Mar-2007, 09:18

Why? It looks like 4x will be usable on G80 for the time-being in DX9 (and perhaps DX10), why wouldn't the R600 with the possibility of literally twice the bandwidth (179+GBPS vs 86.4) if using 2800mhz GDDR4 be an exception for 8x?

I mean granted, new games are taxing, and that might be pushing it to the brink of what to expect, but it is at least possible...especially considering ATi has done a more efficient job with AA/AF in the past using the same amount of BW as nvidia, imagine them having 2x as much. :twisted:

Think this is why nVIDIA came up with what we know as CSAA. 8XQ (8xMSAA) may not be useable for next generation games, but that doesnt mean 8xAA non Q and 16xAA non Q arent useable either. Using Chris Ray's G80 image investigation found on rage3d, you can clearly see that 16xAA non Q has performance hit almost similiar to 8xAA non Q while both are very useable in almost ALL games new (except one or two that i can think of right now) or old.

Im not sure how ATi will approach things, but by the looks of it, instead of what nVIDIA did (with their BW saving technique in the form of CSAA), they may go for the brute force/more traditional approach. (May as well spend the 150gb/s of bandwidth at something).

Would it be possible if ATi implemented SuperAA into single GPUs on the R600?

Tim Murray

03-Mar-2007, 09:46

Would it be possible if ATi implemented SuperAA into single GPUs on the R600?
That doesn't even make sense.

Joshua Luna

03-Mar-2007, 11:08

And the Battlefield Bad Company Demo (http://www.gamershell.com/download_18061.shtml), was that actually run on the R600 or was it run on the Xbox360?US

The recent trailer was on an Xbox 360 dev kit (at ~25fps) according to repi on the B3D console area (http://www.beyond3d.com/forum/showpost.php?p=938202&postcount=10).

vertex_shader

03-Mar-2007, 12:31

The recent trailer was on an Xbox 360 dev kit (at ~25fps) according to repi on the B3D console area (http://www.beyond3d.com/forum/showpost.php?p=938202&postcount=10).

Yes, next week on GDC https://www.cmpevents.com/GD07/a.asp?option=C&V=11&SessID=4555 coming the presentation on R600.

Ailuros

03-Mar-2007, 12:42

Why? It looks like 4x will be usable on G80 for the time-being in DX9 (and perhaps DX10), why wouldn't the R600 with the possibility of literally twice the bandwidth (179+GBPS vs 86.4) if using 2800mhz GDDR4 be an exception for 8x?

If you're that confident that something like 4xAA will be absolutely usable in coming games like Crysis or UT3 (Vista) in resolutions like 1600*1200, then by all means be my guest.

And your assumption about the supposed double bandwidth doesn't hold much merit either, since it takes a tad more than "just" bandwidth to overcome every possible other bottleneck that might occur.

I mean granted, new games are taxing, and that might be pushing it to the brink of what to expect, but it is at least possible...especially considering ATi has done a more efficient job with AA/AF in the past using the same amount of BW as nvidia, imagine them having 2x as much. :twisted:

(Again) if bandwidth alone would be the only defining factor for high resolution AA. But according to your claim above find one of today's taxing games and show me where a R580 can use 6xAA and end up with similar or higher performance than a G71 with just 4xAA. And yes the X1950XTX has a lot more bandwidth than G71 too.

Ailuros

03-Mar-2007, 12:46

That doesn't even make sense.

Why couldn't they add hybrid Super-/Multisampling modes if they'd want to? Granted it wouldn't have to be the exact same algorithm as on Crossfire, but 2xSSAA wouldn't be that much of a problem. Besides I'd be more than glad if they did since it would put some pressure on NV to bring the hybrid modes back for G8x LOL ;)

Geo

03-Mar-2007, 14:05

Why couldn't they add hybrid Super-/Multisampling modes if they'd want to? Granted it wouldn't have to be the exact same algorithm as on Crossfire, but 2xSSAA wouldn't be that much of a problem. Besides I'd be more than glad if they did since it would put some pressure on NV to bring the hybrid modes back for G8x LOL ;)

Of course they could, but why they'd want to call it "Super AA" is pretty mysterious. They've been paying lip service (more so early on, not as much lately) to the possibility of bringing SS back at some point. Always been the generic "We might if customer demand is great enough, but its not a high priority. . . "

Topman

03-Mar-2007, 20:23

Sorry for my poor english... :(

I think SSAA SuperAA(CF) don't work with non stardard framebuffer formats... maybe the 'blend/composition' operation don't work in fp16 applications or even 10:10:10:2 fb format (like ToyShop or Andy shadow demo).

SSAA SuperAA gives horribly results with determined post-effects in applications, like NFS:Carbon, 3dmark05... CF SSAA gives a very strong blur effect (soaped!)... if you disable post-effects, the SSAA works good (with a minimum soft blur side-effect).

SSAA SuperAA works flawlessly in HL2:LC... incredible and perfect IQ...

'Normal' SSAA (like 8XS from GF7) don't give strange results... Works ok (and slow) in any application (but don't work with non stardand framebuffer formats too).

I think R600 has horsepower to make perfect mixed SSAA modes.
[SSAA mode working in SQR(2) X (h X v)]. Maybe using AVIVO engine to make downscaling operation.

bye

SirPauly

03-Mar-2007, 21:18

Twinkie

03-Mar-2007, 21:25

Unknown Soldier

03-Mar-2007, 21:51

Yes, next week on GDC https://www.cmpevents.com/GD07/a.asp?option=C&V=11&SessID=4555 coming the presentation on R600.

Drool

bdotobdot2

03-Mar-2007, 23:06

Yes, next week on GDC https://www.cmpevents.com/GD07/a.asp?option=C&V=11&SessID=4555 coming the presentation on R600.

Hmmm, strange they would use the R600, just cause nvidia has been on the boxes of both BF2 and BF2142. Still be great to see thou!!

Kaotik

03-Mar-2007, 23:50

Hmmm, strange they would use the R600, just cause nvidia has been on the boxes of both BF2 and BF2142. Still be great to see thou!!

And this makes it strange.. how? It's BF"3", not BF2, not the first time company X uses hardware Y for game A and hardware Z for game B

Random Digital

04-Mar-2007, 02:57

Kyle Bennet seems to be absolutely sure about that 300w:
http://www.hardforum.com/showpost.php?p=1030710530&postcount=35
http://www.hardforum.com/showpost.php?p=1030710546&postcount=39
300 watts :shock:

Crossfire would need 600 watts. I hope that's not true. I would bet most people would need a new power supply.

Cuthalu

04-Mar-2007, 03:55

300 watts :shock:

Crossfire would need 600 watts. I hope that's not true. I would bet most people would need a new power supply.

It's very unlikely to be true. Info from AMD's R600 demo event say that it's going to be 200w.

trinibwoy

04-Mar-2007, 03:58

Yeah this power consumption thing is way overhyped and way overblown. They probably wanted more of a cushion than 225W provides for overclocking - I'm sure stock R600's will be perfectly fine with dual 6-pins.

Sound_Card

04-Mar-2007, 05:14

man, nothing new today....:sad:

Shtal

04-Mar-2007, 06:21

I don't think anyone is saying Super-AA literally, meaning identical to how CrossFire offers these modes. What would be nice if single R-600 solutions could have the IQ flexible settings of x8, x10, x12, x14 AA to improve immersion.

I would say anything above IQ x10 AA is plenty enough, because you will not notice any differentness anyway if you playing fast movement games at high resolution.

turtle

04-Mar-2007, 06:39

I would say anything above IQ x10 AA is plenty enough, because you will not notice any differentness anyway if you playing fast movement games at high resolution.

Wow, you said something I actually kind of agree with. :cool: :eek: :twisted:

Shtal

04-Mar-2007, 07:19

Wow, you said something I actually kind of agree with. :cool: :eek: :twisted:

You comparing from the past ? :)

Anarchist4000

04-Mar-2007, 22:59

Given the state of the NV drivers for the G80 and that ATI hasn’t released their hw yet; it’s hard to see how this is really a bad plan. We really want to see final ATI hw and production quality NV and ATI drivers before we ship our DX10 support. Early tests on ATI hw show their geometry shader unit is much more performant than the GS unit on the NV hw. That could influence our feature plan.
http://blogs.msdn.com/ptaylor/default.aspx

Interesting little comment on performance I saw posted.

trinibwoy

05-Mar-2007, 00:59

Oh oh.....so Nvidia is holding back progress again eh :D

INKster

05-Mar-2007, 01:06

Oh oh.....so Nvidia is holding back progress again eh :D

Better to have actual DX10 hardware (good or bad, it's there) than to have vapourware for the next 6 months.

I'm tired of waiting and seeing benchmarks, and counter-benchmarks, and rumours about this and that without the proper follow up.
Either i know something solid about the R600 "light" in two weeks, or i'm going with the 8800 GTS 640MB.
This little NV45 has had it. :wink:

Sound_Card

05-Mar-2007, 01:06

Oh oh.....so Nvidia is holding back progress again eh :D

IMHO, I think Nvidia really does have good HQ performance drivers ready to go and has since Janurary at least. They have to be holding back, because what?... 5 months without any performance drivers... ya right Nvidia.

Razor1

05-Mar-2007, 01:08

Well holding back as in nV might have skimmed on something on the GS side of things, its very possible, just that there is nothing to compare it against so.......

It could also be drivers just have to wait and see.

Arty

05-Mar-2007, 01:32

With 8800 Ultra is rearing up full steam and people concluding it to be a cherry picked 90nm G80 (a 80nm would probably be 8900), I think it limits Nvidia's shock value as opposed to a 80nm G80.

I'm sure the consequence of delaying R600 was obvious to AMD and they still took it. :razz:

Mintmaster

05-Mar-2007, 01:35

Oh oh.....so Nvidia is holding back progress again eh :D
I've been critical about NVidia on this front, especially with regards to dependent texturing and dynamic branching, but software is just waaay behind the hardware at this point.

Moreover, I think the value of the GS is exaggerated. It can help you be more efficient in certain scenarios (especially stencil shadows, but IMO they're obsolete), but there isn't much you can do with a GS that you couldn't do before.

I'm a feature whore, and G80 has everything we need. We should be thankful that they didn't try making a 1.4GHz 48-pipe G70 with DX10 support tacked on.

Andrew Lauritzen

05-Mar-2007, 02:31

I've been critical about NVidia on this front, especially with regards to dependent texturing and dynamic branching, but software is just waaay behind the hardware at this point.
Yeah I was pretty unhappy with the dynamic branching on the G70 and NV40, but in retrospect, it doesn't look like it hurt NVIDIA very much, as we're only now starting to see end-user products that expose the inefficiency... and we have G80 now which solves the problem.

Moreover, I think the value of the GS is exaggerated. It can help you be more efficient in certain scenarios (especially stencil shadows, but IMO they're obsolete), but there isn't much you can do with a GS that you couldn't do before.
Couldn't agree more. To be honest, I'm more excited about stream out (which theoretically can be quite efficient) than geometry shaders.

I'm pretty happy with the G80. Just "making it faster" will keep me happy for quite some time. If the R600 lives up to the G80, I'm a happy camper.

silent_guy

05-Mar-2007, 03:07

Couldn't agree more. To be honest, I'm more excited about stream out (which theoretically can be quite efficient) than geometry shaders.

Quick question: coud you explain in what practical situations one can use stream out? How does this help a game? I though stream out was something that was needed for GS, but from what you wrote here, my understanding of it is clearly too limited.

Razor1

05-Mar-2007, 03:28

SO is used with GS, so things can be saved to a buffer to be reused, I guess you think of it like a render target for geometry. It could also be used for shadowing I think.

Anarchist4000

05-Mar-2007, 04:13

Particles are another feature that comes to mind. It seems best suited for something procedural. Also used on terrain as a sort of geo-mipmapping/adaptive tessellation replacement seems like a distinct probability. For instance if you had a triangle and each vertex had a normal you could create additional triangles to better represent the desired curve in the terrain. Where a vertex shader can only move a vertex around and GS would be able to morph entire shapes. Another example would be turning a pyramid into a perfectly formed half sphere.

If you wanted you could continue breaking down primitives until they were smaller than pixels themselves.

Zengar

05-Mar-2007, 04:42

I've been critical about NVidia on this front, especially with regards to dependent texturing and dynamic branching, but software is just waaay behind the hardware at this point.

Aren't you mixing something up? It's true, NV's branching performance is really bad , but they didn't have texture indirections limits since the FX series. Not that I want to start another flame war here :-) but Nvidia cards were always more more attractive to developers because they always had the newest stuff exposed (but not always at good speed). ATI still has problems with implementing several important OpenGL extensions :-/

IMoreover, I think the value of the GS is exaggerated. It can help you be more efficient in certain scenarios (especially stencil shadows, but IMO they're obsolete), but there isn't much you can do with a GS that you couldn't do before.

I though that calculating the tangent space in the GS would be a nice idea, that will save that annoying per-vertex data that don't belong into vertex declaration ;-)

P.S. A sidenote, Nvidia employee mentioned that their GS was designed to output small numbers of vertices per primitive.

3dcgi

05-Mar-2007, 05:29

Quick question: coud you explain in what practical situations one can use stream out? How does this help a game? I though stream out was something that was needed for GS, but from what you wrote here, my understanding of it is clearly too limited.
Streamout isn't needed for GS to work. It's just a tap in the pipeline where you can write the output of the GS to memory. If you want the output of your VS in memory then you write a passthrough GS that does nothing but streamout. I'm not sure why streamout was tied to the GS, but presumably it was thought to make it easier on the hardware developers to only support it in this case. Saving the result of vertex skinning is one example Microsoft has given.

Ailuros

05-Mar-2007, 05:52

Oh oh.....so Nvidia is holding back progress again eh :D

No it's AMD holding back indirectly a potential driver update for my machine ROFL :D

Ailuros

05-Mar-2007, 05:55

Small trick question for the dev guys amongst us: wouldn't you have preferred a programmable primitive processor instead of "just" the GS?

CarstenS

05-Mar-2007, 09:33

D'you mean R600 is featuring one of those – relics from the original R400-Design, when a PPP was still thought to be required? Cool!

DemoCoder

05-Mar-2007, 10:35

The G80, it appears, doesn't have a "GS" unit. Geometry Shaders are an artifact of the drivers. From the CUDA point of view, I imagine a kernel working on GS inputs, leaving generated vertices in the PDC, and I imagine the G80 DX10 driver essentially converts GS into an underlying equivalent as described above. There might be some special support in the scheduler with respect to tagging GS threads, but then again, it could just be an artifact of the driver using general purpose HW.

Given the size of the PDC, that's probably why GS outputs are limited in size. Perhaps NVidia hasn't written code to write GS-generated vertices to video RAM, or perhaps they did and performance sucked.

In any case, GS to me seems way too limited in many cases, and I bet developers end up using CUDA/Brook/CTM + middleware to do any real geometry amplification tricks. The only thing GS has going for it is simplicity and standardization of the programming model.

Ailuros

05-Mar-2007, 12:53

D'you mean R600 is featuring one of those – relics from the original R400-Design, when a PPP was still thought to be required? Cool!

I'll stick my head out and ask since when anything fixed function is programmable?

CarstenS

05-Mar-2007, 13:19

I'll stick my head out and ask since when anything fixed function is programmable?

*pulls_ailuros'_head_even_further*
R600 has fixed-function units to accelerate GS?
*/pulls_ailuros'_head_even_further*

The G80, it appears, doesn't have a "GS" unit.
How does any specific USA qualify for having a "XX" (PS,VS, GS, WFT-S) unit? From my laymans perspektive, different operations only being "driver artifacts" is what a USA is all about. (?)

vertex_shader

05-Mar-2007, 14:44

Fu(a)dZilla in action :lol:

R600 GDDR4 board postponed

GDDR4 board with 1024 MB of memory got delayed till middle of May. We are talking about the availability of such a product. The launch card is now 512MB GDDR3 card codenamed Catseye, production number 102-B006. We are talking about 9.5 inch board with Dual DVI + Vivo, HDMI and HDTV support.

The launch is scheduled for early May and so is the availability of this GDDR3 card. GDDR4 card will be available later than early May, middle of the month if not later.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=65&Itemid=1)

GDDR4 shortage?

Geo

05-Mar-2007, 15:06

The only thing GS has going for it is simplicity and standardization of the programming model.

Well, that's quite some "only"! I'm not so sure it would be a positive for CUDA and CTM to start impinging heavily on game development. Industry standard APIs are a major goodness. At least most of the professional GPGPU applications being discussed are big dollar opportunities for a relatively small number of apps so could be a little more open to writing specific backends for given GPUs. . . a bit like the professional market.

pelly

05-Mar-2007, 15:12

Interesting info in the Fudzilla forums by Fudo...Sounds like he'll have something posted soon....though who knows how much stock to put in this....Could just be trying to drive traffic to the new website...

http://www.fudzilla.com/forum/index.php?topic=9.0

"GDDR 4 had some problems and got delayed and GDDR3 one wont be fast enough, and there is a driver issue as well. No problems my butt who cancles 200+ journalists event without a massive reason. stay tuned will write about it soon enough. We also have some exclusive OCZ memory to tell you about."

vertex_shader

05-Mar-2007, 15:32

Arty

05-Mar-2007, 15:46

When the GDDR3 version not fast enough, than why AMD want release the GDDR3 card first, and not wait 2 weeks? may is still in Q2, and after so many delay this 2 weeks not really counting.
Not fast enough compared to what? My truck? XTX? :lol:

Anon Lamer

05-Mar-2007, 15:47

When the GDDR3 version not fast enough, than why AMD want release the GDDR3 card first, and not wait 2 weeks? may is still in Q2, and after so many delay this 2 weeks not really counting.

Oh the GDDR4 version will certainly be present at launch, it just wont be present in stores. :wink:

vertex_shader

05-Mar-2007, 15:58

Oh the GDDR4 version will certainly be present at launch, it just wont be present in stores. :wink:
I not want any paper launch :roll:

vertex_shader

05-Mar-2007, 16:01

Not fast enough compared to what? My truck? XTX? :lol:

8800gtx, this is the logic, but Fuad talking in Fu(a)dZilla language, sometimes i can't decode all part :wink:

Sound_Card

05-Mar-2007, 16:02

Oh the GDDR4 version will certainly be present at launch, it just wont be present in stores. :wink:

Because they will be sold out.:razz:

Arty

05-Mar-2007, 16:12

8800gtx, this is the logic, but Fuad talking in Fu(a)dZilla language, sometimes i can't decode all part :wink:
Well according to logic, it means that Fuad is leaving always a way out of his rumors.

3dilettante

05-Mar-2007, 16:16

The performance part of that rumor would need more substantiation.

If the memory interface is as wide as believed, then why wouldn't using GDDR3 equivalent to what's on Nvidia's products still offer a significant bandwidth advantage?

Mintmaster

05-Mar-2007, 16:28

Yeah I was pretty unhappy with the dynamic branching on the G70 and NV40, but in retrospect, it doesn't look like it hurt NVIDIA very much, as we're only now starting to see end-user products that expose the inefficiency... and we have G80 now which solves the problem.Oh, most definately. I think ATI was pretty stupid to devote so much die space to an architecture whose primary advantage is DB speed. To me it's more of a philosophical thing. The biggest feature of PS3.0 is DB, so to cop out on this feature means it's not really PS3.0 and much more like PS2.0a (the same is true with ATI's VS3.0 without VTF). I have yet to see a PS3.0 game today that wouldn't look and perform identically on NV4x/G7x using PS2.0a. It's the FP16 blending that set those cards apart from R4xx.

Aren't you mixing something up? It's true, NV's branching performance is really bad , but they didn't have texture indirections limits since the FX series. Not that I want to start another flame war here :-) but Nvidia cards were always more more attractive to developers because they always had the newest stuff exposed (but not always at good speed). ATI still has problems with implementing several important OpenGL extensions :-/I'm talking about way earlier. EMBM on the Matrox G400 was the first form of dependent texturing, and then ATI introduced it on their Radeon (even the Kyro had it). NVidia not only omitted it on the Geforce, Geforce2, and Geforce4 MX, but I assume their devrel encouraged the use of the DX8 texm3x3vspec instruction for shiny water instead of EMBM/texm2x2*. The latter is much more accurate because it permits reflections/refraction of local geometry. Even in the DX8 generation, NVidia decided to limit the math that a texture lookup was dependent on. PS1.4 made a lot more sense than the fixed instructions of PS1.0-PS1.3. If ATI was smart enough to promote parallax mapping in the R200 days, they would have struck gold.

In terms of being more attractive to developers due to features, you're right beginning with NV40. Before that, however, I strongly disagree. Of course drivers were a different matter, especially OpenGL.

SirPauly

05-Mar-2007, 16:54

I would say anything above IQ x10 AA is plenty enough, because you will not notice any differentness anyway if you playing fast movement games at high resolution.

I thought so as well.............but there is another area: EATM. In the past, x10 AA was my quality setting with performance adaptive( x4 Alpha test). A nice balance of performance and image quality. Since EATM is such a tiny hit - have raised the settings to x14 AA with x12 EATM -- more image quality and performance. But why? The changing of the multi-samping, changes the quality of the EATM -- I may not see the difference over-all with the edge quality --from x8 ( except in absolute worse-case color contrasting areas) but can see the difference from the added EATM.

Already pleased with the R-600 because it will offer EATM. The biggest hit in my gaming was super-sampled with Adaptive. However, this performance features really offers some nice quality considering its a performance feature.

trinibwoy

05-Mar-2007, 17:32

WTH is EATM and is it tangible or is it another phantom like TAA?

trinibwoy

05-Mar-2007, 17:34

"GDDR 4 had some problems and got delayed and GDDR3 one wont be fast enough, and there is a driver issue as well. No problems my butt who cancles 200+ journalists event without a massive reason. stay tuned will write about it soon enough. We also have some exclusive OCZ memory to tell you about."

That's pretty hard to swallow for two reasons - the X1950XTX has been doing just fine on GDDR4 for quite a while now so availability is presumably decent and ATi knows what they're doing with their memory controller.

The other is that the rumoured 512-bit bus should range between dominate and utterly devastate in the bandwidth arena regardless of what type of memory you stick to it.

Any way, score another point for geo who suggested that ATi might be rethinking the cost disadvantage of a gig of expensive memory that may not even benefit performance.

Jawed

05-Mar-2007, 17:36

WTH is EATM and is it tangible or is it another phantom like TAA?
Presumably it's "Enhanced" Alpha to Mask: MSAA for transparent textures, similar to NVidia's transparency AA. As opposed to super-sampled AA for transparent textures.

Jawed

Natoma

05-Mar-2007, 17:40

Any way, score another point for geo who suggested that ATi might be rethinking the cost disadvantage of a gig of expensive memory that may not even benefit performance.

I wouldn't necessarily say that. Think about the performance difference between the 8800 GTX and GTS when dealing with high res/high aa/high af scenarios. The GTS falls off a cliff in comparison.

With G80+ and R600+ generations, we're heading into the territory where 1600x1200 with AA/AF should be the default resolution, much like R300 made 1024x768 gaming with AA/AF default. A drop from 1GB to 512MB would definitely show up in that arena, particularly as we move forward into DX10 native games coming this year.

Sound_Card

05-Mar-2007, 17:50

Tim Sweeny already said that 1gb of ram is a must for full in game settings in UT3. Dropping the memory on R600 down will be a mistake on ATi's part.

trinibwoy

05-Mar-2007, 17:52

Presumably it's "Enhanced" Alpha to Mask: MSAA for transparent textures, similar to NVidia's transparency AA. As opposed to super-sampled AA for transparent textures.

Thanks, guess I'll go look it up.

Bouncing Zabaglione Bros.

05-Mar-2007, 17:56

Any way, score another point for geo who suggested that ATi might be rethinking the cost disadvantage of a gig of expensive memory that may not even benefit performance.

DDR4 with it's lower power requirements and lower heat output is also an issue when trying to keep within high (as opposed to insane) power/heat requirements.

SugarCoat

05-Mar-2007, 17:57

GDDR4 shortage?

A few things sound strange to me. First, shortage? How about the X1950XTXs, they had GDDR4. Second is its a 512-bit bus, so memory speeds shouldnt be impacting performance near enough for there to be this earth shattering difference against the 8800s. If the card needed GDDR4 to compete that would be odd to say the least. If its not fast enough its because of the configuration of the core and the architecture itself which isnt going to change anytime soon. Drivers could be playing a huge part as well though so i dont put much stock into people saying "cant compete" when the cards still under the curtain.

Tim Sweeny already said that 1gb of ram is a must for full in game settings in UT3. Dropping the memory on R600 down will be a mistake on ATi's part.

that was like...4 years ago, literally.

trinibwoy

05-Mar-2007, 18:02

I wouldn't necessarily say that. Think about the performance difference between the 8800 GTX and GTS when dealing with high res/high aa/high af scenarios. The GTS falls off a cliff in comparison.

True but that doesn't necessarily tell us what's going to happen between 512MB and 768MB. If 512MB is good enough for 1920x1200 4xAA I'm sold.

trinibwoy

05-Mar-2007, 18:04

DDR4 with it's lower power requirements and lower heat output is also an issue when trying to keep within high (as opposed to insane) power/heat requirements.

I'm not so sure that those advantages of GDDR4 are really that important - Nvidia claims that it's 1.5GB GDDR3 Quadro peaks at < 180W.

Frank

05-Mar-2007, 18:56

I agree that it certainly looks like there is still a serious problem with the drivers. Simply because they chose to only show their cards do some GPGPU crunching.

And let's face it: there are still serious problems with the 8800 drivers as well, while that is out for half a year by now.

I think, that it isn't easy to map the heritage of the old model to GPGPU. And lack of sufficient on-chip buffer space could be a good explanation, like DemoCoder said.

SugarCoat

05-Mar-2007, 18:59

And let's face it: there are still serious problems with the 8800 drivers as well, while that is out for half a year by now.

4 months....

Frank

05-Mar-2007, 19:39

4 months....
Ok. My mistake.

Anarchist4000

05-Mar-2007, 19:44

I agree that it certainly looks like there is still a serious problem with the drivers. Simply because they chose to only show their cards do some GPGPU crunching.

Actually it's more likely they went with CTM just to break the 1 TFLOP barrier. DX10 likely wouldn't even be efficient enough to do that. I'm not sure you can gather anything about the state of their drivers based on that demo. DX10 wouldn't even have the capability to run the test that they were running.

I'm sure there is a lot of performance tuning to be done but Nvidia's problems were more of dealing with Vista, not their hardware. ATI has already demonstrated that programming for Vista hasn't been much of a problem and I don't anticipate them having difficulty programming for their own card, especially if it's very similar to xenos.

INKster

05-Mar-2007, 19:52

And let's face it: there are still serious problems with the 8800 drivers as well, while that is out for half a year by now.

Problems with the Windows Vista drivers, not the XP ones.
A normal consequence of coding for a very different API, i'd guess...

I'm expecting teething pains for AMD's DX10 drivers too, once R6xx is out.
But perhaps they turn out to be less significant, due to less mediatic exposure (now shared with Nvidia), and with the Vista launch craze now well behind them.

mhouston

05-Mar-2007, 19:53

Their test was written using Brook and the DX9 backend. CTM isn't magic, just like CUDA isn't magic. However, a simple MAD test with only a few registers being used, no texture lookups, and lots of iterations and pixels should run just about any board at full tilt. Where things get more complicated is a mix of instructions which pushes how the compiler and hardware do instruction scheduling.

Razor1

05-Mar-2007, 19:54

turtle

05-Mar-2007, 20:04

unless its just not cost effective right now.

I think that's precisely part of it. Not only the 1GB aspect, but perhaps using the highest speed-bin as well. I still don't think we're positive if it uses .9,.8, or .7ns GDDR4. Remember X1950XTX uses the slowest available GDDR4, and that was probably why it was effective to replace the GDDR3 counterpart with. If it uses 1.4 (2800mhz) GDDR4, it may cost a pretty penny right now, but come a month or two I wouldn't be surprised if Samsung (and/or Hynix) have faster parts available in the channel, driving the slower chips' price lower. ATM Samsung is currently sampling 2ghz (4000mhz) GDDR4, surely they will release 1.6 (3200) and 1.8 (3600) speed bins in-between now and when those are available...I imagine 1.6 is coming relatively soon.

Frank

05-Mar-2007, 20:55

As mhouston said, it's really easy to make a test that crunches away at some random calculation using only a single instruction over and over.

It's much harder putting a good kernel, compiler and library in place that runs anything from 10 years ago to the near future on a completely different architecture.

It's like Cell versus the old and proven single-thread CPU.

If you make all your ALUs independent and generic, you need an OS. And everything that goes with that.

While the hardware might be really swell, it's completely understandable if it's hard to get a single type of application, that uses a bloatware platform and of which no two single apps behave the same and use the same functions extensively to run good enough to beat the competition.

A GPU is quite different than many CPUs + OS that emulate one.

SugarCoat

05-Mar-2007, 21:24

Well they did release their second card the x1800xl a month or so before the xt last gen.

But I don't see why they would have any "problems" with GDDR4, doesn't make much sense, unless its just not cost effective right now.

comically enough i think it was quite literally 2 or 3 weeks, same as what faud said is going to happen this time, though that was for supply reasons.

i do think they should paper launch it anyway with the rest, get the reviewers the cards at least just because of how late they are, though they've problably lost my sale unless they have some really nice features that rival large amounts of AA and nice IQ. First GTX that hits $500 with or without rebate will problably get snatched up by me at this point.

Andrew Lauritzen

05-Mar-2007, 22:25

Quick question: coud you explain in what practical situations one can use stream out? How does this help a game?
As mentioned earlier, stream out is just a way to output post-GS data to a buffer, that can then be fed directly back into the pipeline. An obvious use is to skin meshes once, even if multiple shading passes are required

I guess personally I'd be happier to just have a more programmable pipeline than a new stage, and SO and render-to-VB ("views" in D3D10) are a step in the right direction. ATI's skinning demo using R2VB was a good example. Still, with the hardware unified and the distinction now made only in scheduling, theoretically it will boil down to the same thing.

To be honest though, the only thing that really interests me about GS is flat-out duplicating geometry for stuff like rendering to cube maps or doing jittered super-sampling. A trivial use, definitely, but one that can actually be made efficient by the hardware.

That's possibly a little strong, as there are good uses for the GS (particles, etc). It's just not particularly revolutionary IMHO and thus I'm not overly concerned about the performance.

SirPauly

05-Mar-2007, 22:56

To be honest though, the only thing that really interests me about GS is flat-out duplicating geometry for stuff like rendering to cube maps or doing jittered super-sampling.

I could use a jittered super-sampled on a single chip solution.

Andrew Lauritzen

05-Mar-2007, 23:58

I could use a jittered super-sampled on a single chip solution.
You can do it with DX9 cards pretty easily - it's just that the GS potentially eliminates having to recull, resubmit and retransform the scene, branching nicely just before rasterizing and shading.

clement

06-Mar-2007, 04:33

comically enough i think it was quite literally 2 or 3 weeks, same as what faud said is going to happen this time, though that was for supply reasons.

i do think they should paper launch it anyway with the rest, get the reviewers the cards at least just because of how late they are, though they've problably lost my sale unless they have some really nice features that rival large amounts of AA and nice IQ. First GTX that hits $500 with or without rebate will problably get snatched up by me at this point.

Too bad you didn't keep a look out better, ZipZoomFly had the Foxconn 8800GTX for $507 after MIR (free shipping), but that rebate has since expired (was like that for most of January and February). I came VERY close to pouncing on that one, but I've been waiting for the Core 2s to drop in price before I upgrade (end of April I'm hearing?). By that time R600 should be on the verge of launch (I'm hoping), with real competition in that performance point I'm hoping for a strong(er) $/perf ratio.

Ailuros

06-Mar-2007, 05:58

As mentioned earlier, stream out is just a way to output post-GS data to a buffer, that can then be fed directly back into the pipeline. An obvious use is to skin meshes once, even if multiple shading passes are required

I guess personally I'd be happier to just have a more programmable pipeline than a new stage, and SO and render-to-VB ("views" in D3D10) are a step in the right direction. ATI's skinning demo using R2VB was a good example. Still, with the hardware unified and the distinction now made only in scheduling, theoretically it will boil down to the same thing.

To be honest though, the only thing that really interests me about GS is flat-out duplicating geometry for stuff like rendering to cube maps or doing jittered super-sampling. A trivial use, definitely, but one that can actually be made efficient by the hardware.

That's possibly a little strong, as there are good uses for the GS (particles, etc). It's just not particularly revolutionary IMHO and thus I'm not overly concerned about the performance.

Some will excuse me for repeating the same question: so would you like the idea of a fully programmable primitive processor in the less foreseeable future (beyond D3D10)? I'm merely asking in order to see what developers actually would like to see after D3D10.

3dcgi

06-Mar-2007, 06:36

Some will excuse me for repeating the same question: so would you like the idea of a fully programmable primitive processor in the less foreseeable future (beyond D3D10)? I'm merely asking in order to see what developers actually would like to see after D3D10.
I'm not a game developer so I can't answer your question, but I have a question of my own. How do you define a fully programmable primitive processor or what is DX10 lacking to make it one?

Shtal

06-Mar-2007, 08:00

I thought so as well.............but there is another area: EATM. In the past, x10 AA was my quality setting with performance adaptive( x4 Alpha test). A nice balance of performance and image quality. Since EATM is such a tiny hit - have raised the settings to x14 AA with x12 EATM -- more image quality and performance. But why? The changing of the multi-samping, changes the quality of the EATM -- I may not see the difference over-all with the edge quality --from x8 ( except in absolute worse-case color contrasting areas) but can see the difference from the added EATM.

Already pleased with the R-600 because it will offer EATM. The biggest hit in my gaming was super-sampled with Adaptive. However, this performance features really offers some nice quality considering its a performance feature.

Not many people cares much about EATM anyway.... (Small differentness) - Example: going from 4X Anisotropic to 8X I see reasonable differentness, but going from 8X to 16X I have hard-time notice any differentness at all.

What I really like is Nvidia's Anti-Aliasing X32 AA with their Quad GF-7950GX2 cards.
It looks almost perfect!
http://www.hardware.fr/marc/tridam/quadsli/32xtaa_quad.png
[Edit: But it is well possibility that Dual R600 may have something similar to NVidia's transparency X32 AA]

bigtabs

06-Mar-2007, 12:14

I don't think many people care about EATM due to the difficulty of enabling it whilst it's still 'in beta'. If it were in the CCC then I think alot more folks would be singing it's praises. The performance of it is very good.

32xAA had supersampling in it. I don't think nVidia are going to let us have SS this gen. Probably for marketing reasons (high AA numbers 8x 16x etc being very performant in benchmarks). So if you see 32xAA making an appearance it perhaps won't be even close to the quality of what you've seen before.

ChrisRay

06-Mar-2007, 12:18

I don't think many people care about EATM due to the difficulty of enabling it whilst it's still 'in beta'. If it were in the CCC then I think alot more folks would be singing it's praises. The performance of it is very good.

You won't be seeing anything like 32xAA for a while with nVidia hardware. They don't seem to want us to be using supersampling any more, at least on their 8 series. Any 32x modes, like the 16x and 8x will not be nearly as nice as the old 6 and 7 series versions, as it's all just MS.

The very fact that the Quatro supports it shows that it can be left within the registry to enable. Alot of Quatro modes are usable by Geforce 6/7 hardware.

bigtabs

06-Mar-2007, 12:23

What's that got to do with it being available for the 8 series. I know it's available for 6 + 7?

I realise you are answering my questions in the other thread at rage 3d on this same topic chris, so you might feel I'm on a bit of a mission regarding this subject when you see me commenting on it here too. I'm just rightly pissed as a consumer. I really value SS modes and don't like the fact that they've been removed. I didn't like it when I lost SS modes going from nVidia to ATI. I don't like ATI not including SS for single cards, and I don't like nVidia dropping their support for it too.

I appreciate your input, as ever.

ChrisRay

06-Mar-2007, 13:19

Well. Usually hybrid AA modes get developed for professional hardware first. Then end up being trickled down into the registry of the desktop hardware. However I somewhat misunderstood the post here. So leave it at that. As far as Nvidia giving up on supersampling. I can assure that my conversation with them have not led me to believe that they are "dropping" SS permanently. But its simply not a priority feature for them right now. Things that are certainly coming first will be Vista stability, further XP tweaks, and SLIAA. I am still a little vexed that SLIAA hasnt made it public yet. For the moment best we can do is wait. Believe me when I say I've made sure the guys over at Nvidia have seen the lack of supersampling complaints. ((specially yours :P)) So they know the niche enthusiast wants them back.

Chris

Sound_Card

06-Mar-2007, 16:21

Tommorw is the day where DICE and ATi demo Bad Company on R600.:wink:

Andrew Lauritzen

06-Mar-2007, 17:10

Some will excuse me for repeating the same question: so would you like the idea of a fully programmable primitive processor in the less foreseeable future (beyond D3D10)?
You'll have to explain exactly what that entails to me. Quite honestly what I would like is a bunch of a fast stream processors and a rasterizer (not dissimilar to the G80 hardware, but not exposed like that in software)... the rest can be built efficiently on that base. Oh, and I wouldn't complain about a good chunk of cell-like local memory for the processors :)

On a completely different note, please pardon my ignorance: isn't EATM just alpha-to-coverage? Why is ATI getting so much attention for this when IIRC NVIDIA had it quite a while before ATI (i.e. Geforce 3 era). Has it not been usable (at least in GL) for quite a long time now? Why are we waiting on a new architecture/drivers?

bigtabs

06-Mar-2007, 17:20

I don't mean to herald EATM as the second coming of Jesus or anything. It's just a nice feature that's new to those who currently have an ATI card.

God knows why they're holding on to it, probably to slap an extra entry on to the R600 list of tricks.

vertex_shader

06-Mar-2007, 18:26

Some new Fu(a)dZilla rumor/FUD:
R600 performance questionable
The performance as we learned was too close to G80 and it was even losing in some tests to Nvidia's high end G80. It is nothing to be worry about as the clocks can be adjusted and I can guaranty that R600 will end up faster than G80 but the real trouble is that Nvidia works on the new chip and will be ready to jump on ATI's delayed back. There might be a new driver as well.

As soon as ATI kick the G80 in the nutz Nvidia will be ready with its new chip and will introduce so call Geforce 8900 GTX cards. R600 is definitely one of the biggest misses for ATI in its long history. It will definitely announce the end of an era, filled with delays.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=68&Itemid=1)

And looks like Hector Ruiz not happy anymore about the R600 (Orton days counted in AMD?:smile:)
R600 disappoints Hector
The big boss confirmed that it is bad as the R600 is being late but he also said that it is such an outstanding product that people will understand the potentials of such a product.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=72&Itemid=1)
I belive when i see :wink:

SirPauly

06-Mar-2007, 18:38

Why read Fuad -- when you can listen to it yourself?

http://www.corporate-ir.net/ireye/confLobby.zhtml?ticker=AMD&item_id=1473686

bigtabs

06-Mar-2007, 20:03

Just listened to it - he says exactly:

"
Well, any time you're late with a product you are disappointed, and you do obviously suffer from it. But it's such an outstanding product, you can see when people begin to see the very high end capability of that technology - I think it's going to find a home pretty readily, and we're very bullish about what it can do, and we're pretty excited about what we're going to be able to offer our customers and partners.
"

trinibwoy

06-Mar-2007, 20:32

Doesn't exactly inspire confidence. R600 will have to speak for itself.

zgemboandislic

06-Mar-2007, 23:42

Doesn't exactly inspire confidence. R600 will have to speak for itself.

Yup. I'm tired of all this pre-release talk with words like beast, monster and what not excessively used until the product is released and then that beast/monster always seems to run back to its closet and wait for the next generation. :twisted:

Rangers

07-Mar-2007, 00:45

Some new Fu(a)dZilla rumor/FUD:

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=68&Itemid=1)

And looks like Hector Ruiz not happy anymore about the R600 (Orton days counted in AMD?:smile:)

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=72&Itemid=1)
I belive when i see :wink:

So..we hated Faud so much at INQ now we chase him to his new site in the hunt for more FUD?

Please, let it die.

Ailuros

07-Mar-2007, 05:24

I'm not a game developer so I can't answer your question, but I have a question of my own. How do you define a fully programmable primitive processor or what is DX10 lacking to make it one?

Something like programmable tesselation/HOS for instance. I recall a presentation from NV where the question rose whether "we're there yet" in D3D10 with procedural geometry and the answer was "not quite yet". I don't recall the exact text/phrasing, but I'm sure there were specific geometry related functionalities scratched through early "DX-Next" drafts.

I'm merely wondering what the next step might look like.

vertex_shader

07-Mar-2007, 12:07

R600 to have a new driver
It looks like the big bad R600 had a bad driver. The driver was ok but the performance could be even better. The final goal is clear R600 have to be faster than G80 and a potentially even faster G80 chip.

The 8.351 was supposed to be a launch driver but the plans have change. DAAMIT now has a new launch driver. This one will be known as 8.361 and its first version is scheduled for late march.

The WHQL of the same driver is expected by the middle of April, still on time for the early May announcement. A lot of changes and we hope it will make R600 look good on launch.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=81&Itemid=1)

vertex_shader

07-Mar-2007, 14:53

http://img141.imageshack.us/img141/5471/r6xxgpusxb4.jpg

Left to right: rv410 110nm, r520 90nm, r580 90nm, r600 80nm, rv630 65nm.
Link (http://forums.vr-zone.com/showthread.php?t=133860)

3dilettante

07-Mar-2007, 15:23

That is a big chip.

Interestingly, there are fewer (though larger) resistors on the top of the package compared to the R5xx cores.

RV630 looks almost as clean on top as RV410.

_xxx_

07-Mar-2007, 15:50

Larger resistors are used because of the higher current/power consumption, I guess.

Twinkie

07-Mar-2007, 20:03

So that image of the R600 core was infact not fake as it was assumed quite some time ago when it first popped up in the interweb?

Kaotik

07-Mar-2007, 20:05

So that image of the R600 core was infact not fake as it was assumed quite some time ago when it first popped up in the interweb?

Huh? As far as I can remember, it was judged as legit R600 almost immediately?

Arty

07-Mar-2007, 20:22

So that image of the R600 core was infact not fake as it was assumed quite some time ago when it first popped up in the interweb?
You are confusing it with the image with a 45° rotated die-pin shot.

Twinkie

08-Mar-2007, 04:30

You are confusing it with the image with a 45° rotated die-pin shot.

Oh.. i must be suffering from short term memory loss :lol:

Shtal

08-Mar-2007, 05:21

I was wondering who will win ATI Randeon X2900XTX or upcoming Nvidia 8900GTX.

If you look at Nvidia 8800GTX which has 575MHz GPU-clock and 900MHz-GDDR3 memory. Based on the rumorer information that 8900GTX will be on 80nm tech and be clocked at 700MHz GPU and 1100MHz GDDR4 memory and it will still be at 384bit memory interface. "Plus 25% more shaders on G81"

By just increasing core clock on G80 575Mhz to 700MHz, we are actually talking example like "7800GTX 256MB" against "7900GTX 512MB" similar performance increase.

Since R600 only has 16 ROP's and G80 has 24 ROP's. Giving the fact that high clock speed on 24 ROP's will be significant performance increase.

I was wondering how R600 will compete against G81 Nvidia 8900GTX.

Natoma

08-Mar-2007, 05:29

Where do you get that R600 only has 16 ROPs?

Sobek

08-Mar-2007, 05:35

Where do you get that R600 only has 16 ROPs?

Perhaps someone's been reading a little too much L'inq...

http://www.theinquirer.net/default.aspx?article=35707

Rangers

08-Mar-2007, 05:42

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=81&Itemid=1)

FUDZILLA?

FUDZILLA?

AHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH!!!!!!!!!!!!

Shtal

08-Mar-2007, 05:44

Where do you get that R600 only has 16 ROPs?

32 texture mapping units, 16 raster operation units;
http://www.xbitlabs.com/news/video/display/20070103072506.html

Ailuros

08-Mar-2007, 05:58

Where do you get that R600 only has 16 ROPs?

Granted nothing's definite until the launch, but would it be really that irrational? If I'd assume that those rumoured 64 ALUs are arranged in 4 clusters (4*16), what speaks against hypothetical 4 ROP partitions with 4 ROPs and 128bits each? Logically those ROPs should be 2*R5x0 anyway.

Cuthalu

08-Mar-2007, 06:00

http://www.xbitlabs.com/news/video/display/20070103072506.html

That's the level505 fake. Already mentioned thousands of times here and elswhere... :roll:

Shtal

08-Mar-2007, 06:41

That's the level505 fake. Already mentioned thousands of times here and elswhere... :roll:

Fake of NOT fake, I'm not relying on level505 or Inq....
Would you please show me the truth :)

Sobek

08-Mar-2007, 06:47

Fake of NOT fake, I'm not relying on level505 or Inq....
Would you please show me the truth :)

Dude... The truth isn't out there. :razz:

Shtal

08-Mar-2007, 06:55

Dude... The truth isn't out there. :razz:

I would like to add that the truth about the delay of R600 several times, soon to be revealed also.... :)

Sobek

08-Mar-2007, 07:43

I would like to add that the truth about the delay of R600 several times, soon to be revealed also.... :)

Well then...this is a season finale to tune into!

AlexV

08-Mar-2007, 09:19

No.

Unknown Soldier

08-Mar-2007, 11:35

Dave wannabes

Ollo

08-Mar-2007, 12:15

Yes.

Love_In_Rio

08-Mar-2007, 12:49

already posted ? R600 80 4D ( "320 multiply-accumulate units" ) ALUs ?

http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=197700269

vertex_shader

08-Mar-2007, 12:59

A13 is the R600 final silicon

We got news that DAAMIT actually won't change the R600 silicon. R600 will stay the same and the driver, a PCB design and a clock were the factor that pushed the launch.

The chip is fine, at least one good news from DAAMIT. But still the fans won't benefit from it as the cards got delayed to some point in May.

The chips will be clocked differently than the originally planned but clocks won't be announced until some point in April. The beta cards are running, with an NDA you might actually see one sooner than expected.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=96&Itemid=1)

Tim Murray

08-Mar-2007, 15:25

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=96&Itemid=1)
Duuuuuuuuh.

Bouncing Zabaglione Bros.

08-Mar-2007, 16:06

_xxx_

08-Mar-2007, 16:24

BZB, it's the fun factor methinks...

3dilettante

08-Mar-2007, 16:24

neliz

08-Mar-2007, 16:34

it's always in the but..

R600 to have ZOMG pipes!

We have just confirmed that R600 has ZOMG pipes! from our sources, running at wtfsauceGHz but it could be more or less when the product finally launches.
We heard it will beat G80 but it will lose some..

3dilettante

08-Mar-2007, 16:40

Predication is great.
When there are two possible paths, execute both and then conveniently forget about the one that turns out to evaluate as false.

It's after-the-fact journalism, or being a historian without the history.

Fuad's got a tough job. He has to play the role of two completely opposed historians ahead of time.

zgemboandislic

08-Mar-2007, 17:16

Right or wrong, he obviously has a lot of readers, and probably makes enough cash. It's what it's about after all, isn't it?

trinibwoy

08-Mar-2007, 17:47

I really, really, really hope R600 is tangibly faster than G80. I don't want some Nvidian WaltC rising out of the mist and bashing us over the head about this for the next 4 years :lol:

icecold1983

08-Mar-2007, 18:44

i hope its tangibly faster so i can have a better video card....

_xxx_

08-Mar-2007, 19:51

I really, really, really hope R600 is tangibly faster than G80. I don't want some Nvidian WaltC rising out of the mist and bashing us over the head about this for the next 4 years :lol:

I predict a desaster for ATI this year based on my predictions of nV strategy. I'm pretty confident that they could release the G81 or whatever it turns out being within 2-3 months from today and maybe they already even have the successor as good as ready.

I still predict R600 being 15-20% faster overall and nV releasing their refresh/successor shortly after if not simultaneously.

But what will kill it (or has already) is that people turned away from them and towards nV and that's not likely to change overnight. The only possibility would be to offer it at the same price as the 15-20% slower competition or such actions which are bad for margins and the company in the long run.. Conclusion: either no more high-end stuff but moving towards integration completely or the downward spiral.

The ONLY thing that could save them would be a totally superior product which will stomp the competition with at least +60-70% while NOT costing much more then the card compared. Still, I think the people here still waiting for some AMD miracle should finally wake up. My honest opinion.

I hope you guys don't get this as some sort of negative criticism, this is just what I see when I look at it purely from the business point of view. If they make it, all power to them :) But I'll still buy the most compelling product for me, which is whatever has a healthy balance between price/performance, size and power requirements.

_xxx_

08-Mar-2007, 19:55

Or, to put it short: either R600 will truly rock, or we're all but done with the high-end GFX part of AMD.

trinibwoy

08-Mar-2007, 21:14

i hope its tangibly faster so i can have a better video card....

Why? You're obviously not happy with your 8800 so I doubt a faster card would make you any happier.

Sound_Card

08-Mar-2007, 21:25

Or, to put it short: either R600 will truly rock, or we're all but done with the high-end GFX part of AMD.

That's hog wash.

EVEN if R600 was slower than g80, I doubt that would be the end of it. What % of sales to the high end bring in? More importantly, what is their profit margin? The true purpose of "high end" is to demo your technolgy, show the strengh of your products. Regardless of R600 being late, I doubt that is of the worries AMD faces as a whole. Point being that AMD is still not late to the party in the low end or mid range products. Taking a look at AMD's Q4 results, they went up almost through out the board. So obviously they must be doing something right. I have come to the conclusion that their new AGP high end parts must be selling like hot cakes that along with their x1950pro, x1950gt, and x1950xt. Anyway's Q1 results for this fiscal year should prove intresting in just how much R600 could be hurting AMD. Notice the under statement?:razz:

Razor1

08-Mar-2007, 21:57

hmm interesting AMD didn't reach thier expected gross revenue Q1 :oops: , sure you want to make that statement sound so good for AMD?

Geeforcer

08-Mar-2007, 22:28

If we learned anything from NV30 is that a company can err badly and still pull through and blossom - and ATI, being a part of AMD now, is in far better position than Nvidia was. I am not saying that R600 = NV30 II: The Shadering, just that if latter didn't kill Nvidia I can't see how whatever happend to former will signify the end of ATI highend.

Sound_Card

08-Mar-2007, 22:29

hmm interesting AMD didn't reach thier expected gross revenue Q1 :oops: , sure you want to make that statement sound so good for AMD?

I think your confusing with their CPU's and their drastic price cuts. Not their GPU's.

Razor1

08-Mar-2007, 22:30

I think your confusing with their CPU's and their drastic price cuts. Not their GPU's.

its all one and the same now, thier estimates are based on every line they have not just one or the other products :wink:

Sound_Card

08-Mar-2007, 22:34

its all one and the same now, thier estimates are based on every line they have not just one or the other products :wink:

Yes, but if you break it down and compare their market share across the board and their gross/net profits, I'm thinking their "multi media" department is not exactly feeling "doomed".

Sound_Card

08-Mar-2007, 22:36

It seems that people are confusing AMD's "doom" with R600. Much of AMD's sucess and future will be based on how well K10 peforms. R600 is just the cherry on the cake.

Razor1

08-Mar-2007, 22:44

There is no such thing as cherry on the cake in this situation.

Well first thing would be a major restucturing of thier graphics division, next thing would be doomed, and this will have alot to do with how the entire line up of the r6x0 line does at the market place.

Breakdown wise I think they lost some market share in the CPU division and lower margins and of course the GPU division isn't going to make up that loss even if it increased marketshare and increased margins, which neither of those did happen in the GPU division.

Rangers

08-Mar-2007, 22:54

Well if AMD keeps raping Intel in market share like they have been quarter after quarter and not stopping, I dont think they're too worried about it. And that's not even with Barcelona.

Too be honest I see more signs Nvidia is leaving the high end, it can be argued they already have, they are sort of a consistent performance 2nd place now.

The real positive for AMD is being able to bundle CPU/GPU to OEM's, which we haven't even seen the fruits of this merger yet, which could be REALLY dangerous for the other players out there. Intel may have to buy Nvidia soon. Of course, Intel has it's own gfx plans, which could leave Nvidia in a real cold place, which is probably why Nvidia is rushing to develop CPU's I just wonder, too little too late..

Faud..faud predicts everything related to a GPU sooner or later.. Remember his ridiculous R520 predictions...which pretty much ran every gamut of possible configurations at one time or another...

We really need a better industry yellow journalist, you know? I mean, as bad as Faud's info is, NOBODY must talk to him from the inside. As long as he's in the business, how can he have no contacts? Plus, he writes in the most tortured, hilariously bad prose ever..even if he had some actual info once in a blue moon, I dont think he has enough literary ability to convey it.

It'd be really nice to have a SOLID, reputable source for such things..

Tim Murray

08-Mar-2007, 23:03

Razor1

08-Mar-2007, 23:08

Arty

08-Mar-2007, 23:09

I think your confusing with their CPU's and their drastic price cuts. Not their GPU's.
The recent warning was re channel shortages for ... CPUs. There wasnt even one question related to GPU side of things, understandably.

Rangers

08-Mar-2007, 23:09

NVIDIA is a GPU company. AMD isn't. AMD can pull out of the high-end market and survive as a company.

Err, Nvidia could too...

Nvidia have much more business in the low end, and in fact really seem to be moving more that direction and into handhelds.

As well as developing a CPU/GPU hybrid..chipsets etc.

Rangers

08-Mar-2007, 23:10

Why not check any of the financial sites about the possible market share loss, alot of em have been talking about that for this past week ;)

Got any links?

The last stuff I saw was for Q4 06, where AMD continued picking up share, even as you talked about Intel "raping" them.

Razor1

08-Mar-2007, 23:12

Got any links?

The last stuff I saw was for Q4 06, where AMD continued picking up share, even as you talked about Intel "raping" them.

Look around its pretty plan in sight, I mentioned it earlier this week with a link, in the AMD financial thread, and check today at yahoo news if you want to. AMD already warned they weren't going to make expectations and the reasons were marketshare loss to Intel, and lower margins.

trinibwoy

08-Mar-2007, 23:16

Nvidia have much more business in the low end, and in fact really seem to be moving more that direction and into handhelds.

I believe the handheld market is a relatively high margin business. Am I wrong?

Rangers

08-Mar-2007, 23:18

Look around its pretty plan in sight, I mentioned it earlier this week with a link, in the AMD financial thread, and check today at yahoo news if you want to. AMD already warned they weren't going to make expectations and the reasons were marketshare loss to Intel, and lower margins.

Meh, they can lose a bit with Barcelona around the corner..their gains have certainly been relentless over the last 12 qaurters.

Arun

08-Mar-2007, 23:22

Nvidia have much more business in the low endNo, ATI and AMD both have more business in the low end than NVIDIA and Intel respectively.
and in fact really seem to be moving more that direction and into handhelds.I sure as hell hope you're just being vague and didn't mean what I think you meant, because defining handhelds as "low-end" is rather ridiculous.
The last stuff I saw was for Q4 06, where AMD continued picking up share, even as you talked about Intel "raping" them.http://www.chron.com/disp/story.mpl/ap/fn/4613404.html
Please note that this analyst is being more pessimistic about AMD than most, however, afaik.

But let's get back on topic...

Razor1

08-Mar-2007, 23:23

Meh, they can lose a bit with Barcelona around the corner..their gains have certainly been relentless over the last 12 qaurters.

That is just the first sign of Intel's core duo CPU's, it took a few quarters to show up, Barcelona won't be top dog with Penryn coming out in Q3 of this year. That gives 1 quarter for AMD to take Barcelona to the top, thats not enough time to show market impact of Barcelona

Razor1

08-Mar-2007, 23:24

I believe the handheld market is a relatively high margin business. Am I wrong?

Yes it is.

Rangers

08-Mar-2007, 23:27

That is just the first sign of Intel's core duo CPU's, it took a few quarters to show up, Barcelona won't be top dog with Penryn coming out in Q3 of this year. That gives 1 quarter for AMD to take Barcelona to the top, thats not enough time to show market impact of Barcelona

How do you know Penryn will beat Barcelona?

I dont know why you want AMD to die so bad. It's just going to mean one company dominance, and that company will be Intel (in the long run), not Nvidia.

I think AMD should not have bought ATI because that is when everybody started hating them..because everybody hated ATI..

I think IBM needs to buy AMD to ensure they survive :) It sounds like they will go out of business soon from the financial reports otherwise.

Great move, that ATI purchase..

Natoma

08-Mar-2007, 23:35

it's always in the but.. Originally Posted by fud
R600 to have ZOMG pipes!

We have just confirmed that R600 has ZOMG pipes! from our sources, running at wtfsauceGHz but it could be more or less when the product finally launches.
We heard it will beat G80 but it will lose some..

:shock: :lol:

Oh that gets a pos rep. :grin:

Geeforcer

08-Mar-2007, 23:39

Anarchist4000

08-Mar-2007, 23:51

I think the timing of the ATI buyout was more out of necessity than looking for a good buy. They needed to get started designing the 'Fusion' bound parts. The sooner they started the better. Their execution has definitely left something to be desired and I'm sure it's aggravating to a lot of people.

icecold1983

08-Mar-2007, 23:51

imo amd shouldnt have bought ati because they have enough trouble beating intel devoting 100% of their efforts to cpus, buying a gpu company thats been in a distant second for years and having to focus on 2 markets doesnt seem wise to me

Razor1

08-Mar-2007, 23:53

How do you know Penryn will beat Barcelona?

I dont know why you want AMD to die so bad. It's just going to mean one company dominance, and that company will be Intel (in the long run), not Nvidia.

I think AMD should not have bought ATI because that is when everybody started hating them..because everybody hated ATI..

I think IBM needs to buy AMD to ensure they survive :) It sounds like they will go out of business soon from the financial reports otherwise.

Great move, that ATI purchase..

AMD if they bought ATi when they had no problems on the CPU side of things it wouldn't look this bleak, they put themselves at a huge risk because of the buyout, the first sign was thier credit being slashed, ever go to a bank and get a loan denied or people who have to pay increased interest rates on loans. This is a similiar situation. And as margins decrease it only gets worse. AMD expects/expected ATi's profitability to remain similiar over the past 2 Q's and the next few, just like before the buyout, unfortunately ATi lost alot in the past 2 Q's and doesn't look like that will change.

Why wouldn't Penryn be competitive with Barcelona? I don't see Intel playing dead after P4 do you?

Arun

08-Mar-2007, 23:57

I dont know why you want AMD to die so bad. It's just going to mean one company dominance, and that company will be Intel (in the long run), not Nvidia.The goal when predicting the future is not to try to predict what's best for yourself, or for the industry in general. The goal is to predict what is most likely to happen.
I think AMD should not have bought ATI because that is when everybody started hating them..because everybody hated ATI..Either that, or it's also just about the time Intel released Conroe! That likely had a much larger impact on public perception than the ATI acquisition ever could.
I think IBM needs to buy AMD to ensure they survive :)While I don't think IBM is the most likely candidate, I agree that AMD has plenty of engineering talent and IP, which is why no matter what happens, their technology won't die overnight. No matter what happens to the company. Also, while I think the tales of doom have been largely exagerated (especially so timeframe-wise), I fear they are also far from being completely without basis.

Kaotik

09-Mar-2007, 01:04

It seems that people are confusing AMD's "doom" with R600. Much of AMD's sucess and future will be based on how well K10 peforms. R600 is just the cherry on the cake.

Hold your horses there, K10? How about seeing how K8L does first, and then think about the future products.
As far as I know, the thing saying K8L = Turion, K9 = A64 X2, K10 = Barcelona etc was shot down pretty quickly, as K8L was titled as "upcoming product" while Turion was already out, and K8's were designed with dualcores in mind since the beginning

rwolf

09-Mar-2007, 01:51

No, ATI and AMD both have more business in the low end than NVIDIA and Intel respectively.
I sure as hell hope you're just being vague and didn't mean what I think you meant, because defining handhelds as "low-end" is rather ridiculous.
http://www.chron.com/disp/story.mpl/ap/fn/4613404.html
Please note that this analyst is being more pessimistic about AMD than most, however, afaik.

But let's get back on topic...

ATI has more business then Intel in the low end? How is that when Intel has 37-40% market share?

Razor1

09-Mar-2007, 02:28

http://64.233.179.104/translate_c?hl=en&u=http://it.enorth.com.cn/system/2007/02/26/001557853.shtml&prev=/search%3Fq%3Dr600%2B08-03-07%26hl%3Den%26lr%3Dlang_zh-CN%26as_qdr%3Dall

its a bit old so it might have been posted already don't know.

trinibwoy

09-Mar-2007, 02:40

Wow, google did a pretty damn good job of translating that page. Nothing we havent heard before though.

Razor1

09-Mar-2007, 02:47

heh looks like clicking the link doesn't work, copying and pasting works though, wierd. ah sorry well I'll go back into my hole for tonight :grin:

Sound_Card

09-Mar-2007, 05:13

Hold your horses there, K10? How about seeing how K8L does first, and then think about the future products.
As far as I know, the thing saying K8L = Turion, K9 = A64 X2, K10 = Barcelona etc was shot down pretty quickly, as K8L was titled as "upcoming product" while Turion was already out, and K8's were designed with dualcores in mind since the beginning

Weird. I keep seeing sites say K10 now. Something is going on... anyway's my bad if it is incorrect.

Russell

09-Mar-2007, 07:13

Weird. I keep seeing sites say K10 now. Something is going on... anyway's my bad if it is incorrect.

I think the K10 thing is correct. K8L was a name the Inq pretty much made up anyways, so there's no sense of validity there :P

Kaotik

09-Mar-2007, 11:47

AMD executive VP Henri Richard's June 2006 interview with DigiTimes comments on the future processor development:

“ Q: What is your broad perspective on the development of AMD processor technology over the next three to four years?
A: Well, as Dirk Meyer commented at our analysts meeting, we're not standing still. We've talked about the refresh of the current K8 architecture that will come in '07, with significant improvements in many different areas of the processor, including integer performance, floating point performance, memory bandwidth, interconnections and so on. You know that platform still has a lot of legs under it, but of course we're not standing still, and there's a next-generation core that's being worked on. I can't give you more details right now, but I think that what's important is that we're establishing clearly that this is a two-horse race. And as you would expect in a race, sometimes, when one horse is a little bit in front of the other, it reverses the situation. But what's important is that it is a race.

But enough with the offtopic

vertex_shader

09-Mar-2007, 13:20

I hoped we were going to see less of Faud's random ramblings once he left TheInq, but it seems that people have actually followed him to his new site and are still quoting his same old scattergun approach to journalism.

C'mon guys, I know that even a stopped clock tells the correct time twice a day - but that still means it's wrong the other 1438 minutes.

This is a rumor tread, rumor can be turn as FUD too, who knows before (only Ruiz, Orton and some more AMD people :smile: ) , its fun anyway for me, i never said take what he write serious, but many times he has right, so he not spread FUD only :wink:

Here is the latest one :smile:
DAAMIT kills the R600XT

ATI won’t have the usual XTX and the XT alignment. The company decided to change its original plan. At first DAAMIT planned to play the usual game and introduce R600 XTX and R600XT, were the second one is a cheaper and slower clocked versions of the chip.

That won't be the case this time as ATI plans to unify the R600 brand and sell it under one umbrella. The only difference will be in the quantity of memory. The first one out will be GDDR 3 based board with 512 MB of memory codenamed UFO and after it, ATI will release GDDR 4 card with 1024 MB card codenamed Dragonshead2.

One chip only means that the memory will play the difference as well as the memory speed. ATI obviously knows more than it wants to tell the world, and I would bet that the answer is in performance of the R600.

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=106&Itemid=1)

_xxx_

09-Mar-2007, 13:27

Hmm... that would imply that the XT performance sucked or what?

Slappi

09-Mar-2007, 13:33

Fuad is wrong as usual. His site won't last long.

trinibwoy

09-Mar-2007, 13:35

That would be doubtful unless they're expecting something faster than the 8800GTS from Nvidia. There is a massive gap between the GTS and GTX so if R600 is competitive with the GTX then I don't see how they couldn't do an XT that's competitive with the GTS.

It's all very curious. They need a card in the $400-$500 bracket so I don't see how they can forego and XT.

neliz

09-Mar-2007, 13:38

Hmm... that would imply that the XT performance sucked or what?

Vice versa, ATI was the embodiment of GDDR4 so there must be a reason that they can afford to go out with GDDR3 at first and still satisfy their target markets.
unless the r600 stinks and they're just trying to get the boards out as cheap as possible.

so the 512MB G3 will combat the 8800GTX and a month later the 1GB G4 will come to compete with the 8800Ultra.

so that leaves us... with either a cheap R600 or a fast R600.

Arty

09-Mar-2007, 13:41

trinibwoy

09-Mar-2007, 14:13

There's an XL board. :wink:

Doesn't matter what it's called as long as it exists :)

Geo

09-Mar-2007, 14:15

INKster

09-Mar-2007, 15:18

"One chip only. . . "? Am I allowed to cry tears of frustration over how ignorant he is after umpty years of reporting on this stuff?

Is what he's really trying to say is the XT and XTX will have their cores clocked the same (this would be news, as hkepc suggested 10-15% less for XT) and XT won't have units disabled to bump up yields (the hkepc thing implied it wouldn't, so I wouldn't call that part news)?

There surely is a yawning gap in the 8800 family for a price/performance monster to clean up, as my sig has been aimed at.

Geo, the QuadroFX 4600 might give us a hint of what's to come...

As for Fuad, well..., his "journalism" hasn't changed, so we should stop giving him this kind of significance.
Visiting his site is an encouragement for him to continue his mindless antics.

Ollo

09-Mar-2007, 15:27

This is a rumor tread, rumor can be turn as FUD too, who knows before (only Ruiz, Orton and some more AMD people :smile: ) , its fun anyway for me, i never said take what he write serious, but many times he has right, so he not spread FUD only :wink:

Here is the latest one :smile:

Link (http://www.fudzilla.com/index.php?option=com_content&task=view&id=106&Itemid=1)

Perhaps Fuad is a registered user on this board - who knows? Would seem like a good idea if he wants to pimp his new site...

Jawed

09-Mar-2007, 15:35

If the only thing that differentiates "XT" and "XTX" is the speed (and amount) of memory, then I guess that might mean that performance at "XTX" clocks is still seriously bandwidth-limited...

Pretty strange that the "XT" doesn't have lower core clock though, just for yield

With the way ATI mucked about with X1900GT core/memory clocks last year, who knows...

Jawed

trinibwoy

09-Mar-2007, 16:06

If anything ATi puts out is seriously bandwidth limited with > 115GB/s of bandwidth then they are about to bring some serious pain. But then that would mean the XT tramples the 8800GTX since it seems quite happy with only 86GB/s. Why exactly would ATi release a more costly card on top of that when there's a limit to what they can charge for it? I'm sure the XT could command a $600-$650 price tag on its merits.

Jawed

09-Mar-2007, 16:18

If anything ATi puts out is seriously bandwidth limited with > 115GB/s of bandwidth then they are about to bring some serious pain.
Even with 115GB/s AMD should bring serious pain. After all, GTX brings serious pain to X1950XTX in selected games, even if it's not always twice as fast in high-IQ (i.e. high-bandwidth) modes...

But then that would mean the XT tramples the 8800GTX since it seems quite happy with only 86GB/s. Why exactly would ATi release a more costly card on top of that when there's a limit to what they can charge for it? I'm sure the XT could command a $600-$650 price tag on its merits.
Why should NVidia be the only company allowed to push the limits on single-card pricing?

Anyway, my point is that with the shenanigans pulled last year, there's a low limit on the usefulness of such a rumour, regardless of who posts it.

And, for example, X1950XTX isn't really worth the extra money for its extra bandwidth - apparently its real value derives solely from its quieter cooler.

Jawed

trinibwoy

09-Mar-2007, 16:37

Even with 115GB/s AMD should bring serious pain. After all, GTX brings serious pain to X1950XTX in selected games, even if it's not always twice as fast in high-IQ (i.e. high-bandwidth) modes...

Exactly, that's why I'm pondering the need for a GDDR4 version.

Why should NVidia be the only company allowed to push the limits on single-card pricing?

Well, I don't think they've been doing it any more than ATI has. The release of the 8800 series has not affected the high-end price ceiling at all. In fact, it's served to drive very powerful hardware (X1950XTX) into relatively low price levels.

Jawed

09-Mar-2007, 17:09

Exactly, that's why I'm pondering the need for a GDDR4 version.
Huh? If more bandwidth = more performance, then why not?

What we don't know is if ATI designed something with decent bandwidth efficiency.

It'd be foolish to assume that ATI aimed for something less than 2x as fast as R580. The question is, can it maintain that 2x performance margin in high-IQ scenarios?

I reckon it will:

ATI has a fairly good history since R300 of optimising the marriage twixt architecture and available memory-chips
a bevy of patent applications that point towards increasing bandwidth-efficiencyThis recent patent application:

METHOD AND APPARATUS FOR DATA TRANSFER (http://v3.espacenet.com/textdoc?DB=EPODOC&IDX=US2007028029&F=0)

is interesting, because it reduces the turn-around latency associated with DDR, which helps maximise bandwidth utilisation. But I don't know if it's possible to apply this with GDDR4 (I'm dubious, frankly) or whether it's only relevant to GDDR5...

Jawed

trinibwoy

09-Mar-2007, 17:35

Huh? If more bandwidth = more performance, then why not?

More of anything is always = more performance. Simply more performance isn't a primary goal IMO. Getting competitive and profitable products to market is the end goal. If the GDDR3 version is already fast enough to convincingly trounce G80 then there is no incentive to produce a more expensive GDDR4 part.

If it is not fast enough to convincingly trounce G80 with 115GB/s of bandwidth I fail to see how an additional 512GB of RAM and 25GB/s of bandwidth is going to resolve that since it will obviously be lacking elsewhere.

Now, if you would like to say that R600 is bandwidth limited at 115GB/s (amazing if true) then they surely could have just produced a cheaper/smaller/slower chip and brought it to market earlier, no? And I agree that they're probably aiming to at least double R580 performance but I seriously doubt that requires double the bandwidth.

This scenario would result in the same high ASP and a much lower cost. Gross profit goes up, their technical and management ability is well represented and they have performance in the bag for the refresh. I guess what I'm saying is that I don't see ATi going so balls out if they didn't need to.

Frank

09-Mar-2007, 17:55

Fuad is wrong as usual. His site won't last long.
Oh, it will. There is nothing better than a steady stream of pseudo-news to keep interest. Look at this thread for a nice example. ;)

Jawed

09-Mar-2007, 17:59

I guess what I'm saying is that I don't see ATi going so balls out if they didn't need to.
Generational transitions happen fairly rarely (R300, R520) and contain forward-looking tech. Who knows what they're targetting? i.e. there's a wealth of D3D10-related stuff in R600 that we don't know the first thing about. e.g. streamout might be extremely bandwidth-sensitive, more-so than ROPs.

You might just as well ask, why didn't ATI introduce "slow" dynamic branching, like NVidia did. Smaller chip, more profits, la-di-dah. Or slow vertex texturing instead of none, etc., the list goes on. It's not just about specific-chip profitability but marketing and politics...

Having said all that, it's pretty clear they've fucked up on R600's schedule. The later they leave its eventual launch, the better it has to be (especially as its very likely to end up facing a competitor's refresh, doh). Unless its own refresh follows within 3 months or so :lol: ...

Jawed

Razor1

09-Mar-2007, 19:00

Generational transitions happen fairly rarely (R300, R520) and contain forward-looking tech. Who knows what they're targetting? i.e. there's a wealth of D3D10-related stuff in R600 that we don't know the first thing about. e.g. streamout might be extremely bandwidth-sensitive, more-so than ROPs.

Why would you think stream out might be extensively bandwidth sensitive, I think it was already talked about the uses of stream out well immidiate uses, do those situations sound to be bandwidth intensive?

check page 30

http://ati.amd.com/developer/gdc/2006/GDC06-Advanced_D3D_Tutorial_Day-Riguer-Bridging_DX10_and_DX9.pdf

Jawed

09-Mar-2007, 19:34

Why would you think stream out might be extensively bandwidth sensitive [...]?
Because you don't stream data out unless you're going to read it back in again (perhaps not all of it, though?).

And because when you stream data out, any delay in being able to start reading that data back in again slows down the entire process that's dependent upon the streamed-out result.

i.e. streamout isn't an end in itself, its latency will have a knock-on effect on rendering steps that depend upon its results. Theoretically the latency can be reduced by streaming-out in batches (rather than the streamout phase of rendering being programmed as one long-winded batch), but that only works if the GPU can consume SO-batch 1 while SO-batch 3 is being computed and SO-batch 2 is being written to memory. As far as I can tell Xenos is quite happy doing this, but I'm doubtful D3D10 GPUs can...

Regardless, you can't consume SO data faster than you produce it (except instantaneously, of course).

No matter what, the faster you write SO data, the sooner you can consume it all. What we don't know, clearly, is what kind of effect bandwidth has on SO, or the kind of first-generation D3D10 apps that use SO.

But it's worth noting that the output capabilities of the GS unit have been set "low" (4096 bytes of output data per input triangle, but 2048 bytes, I think, streamed out), in order not to push the hardware too much, either with internal post-transform cache or SO tiling/bandwidth usage...

Jawed

trinibwoy

09-Mar-2007, 19:47

Generational transitions happen fairly rarely (R300, R520) and contain forward-looking tech.

Agreed but I'm not talking architecture here. I'm assuming that no matter what R600 is it is going to scale with bandwidth so I don't see the need to start high just for shits and giggles.

You might just as well ask, why didn't ATI introduce "slow" dynamic branching, like NVidia did. Smaller chip, more profits, la-di-dah. Or slow vertex texturing instead of none, etc., the list goes on. It's not just about specific-chip profitability but marketing and politics...

Yeah definitely. That sure worked out well for NV. Maybe I'm just not used to seeing triple digit bandwidth numbers as yet but I have a hard time understanding why or how R600 would possibly use more than 115GB/s.

Jawed

09-Mar-2007, 21:10

Agreed but I'm not talking architecture here. I'm assume that no matter what R600 is it is going to scale with bandwidth so I don't see the need to start high just for shits and giggles.
Maybe it's a matter of SO working versus not working. You know, a bit like AA, back in the day...

Jawed

Arty

09-Mar-2007, 21:14

SO? (Yeah, I'm not quick enough :oops: )

BRiT

09-Mar-2007, 21:25

SO? (Yeah, I'm not quick enough :oops: )

StreamOut.

Arty

09-Mar-2007, 21:35

StreamOut.
http://img144.imageshack.us/img144/1176/dohwe0.gif

jamesnmandy

09-Mar-2007, 21:46

i know one thing, speculating was fun about a year ago or however long we have been hearing about this non-existent product, "DAAMIT" needs to get it together already or find another part of the business to funnel all this time and money into, i wouldn't be surprised if we hear R600 has been scrapped, skipping an entire product cycle, i wouldn't even embarrass myself by releasing an unimpressive product so late

Razor1

09-Mar-2007, 22:00

Maybe it's a matter of SO working versus not working. You know, a bit like AA, back in the day...

Jawed

Jawed ask yourself how much bandwidth does a list of vertices take?

Mintmaster

09-Mar-2007, 22:01

No matter what, the faster you write SO data, the sooner you can consume it all. What we don't know, clearly, is what kind of effect bandwidth has on SO, or the kind of first-generation D3D10 apps that use SO.SO is not going to be very bandwidth intensive, IMO. GF8800, for example, has 2 bytes of BW per scalar shader cycle per clock. Even a simple 4D vector transform would let you sustain 32 bytes per pixel without being BW limited, so for any reasonable SO shader you have gobs of BW.

I understand what you're saying about reading the data back in (possibly multiple times), but it would take quite an extreme (and maybe poorly written) workload to make vertex BW jump so dramatically.

Jawed

09-Mar-2007, 22:59

When SO is used on the back end of a data-amplifying render pass you might have 20-40 bytes in (per triangle) but 500-2000 bytes out (I'm unclear on the maximum amount of data SO can write per triangle).

Not to mention all the bandwidth consumed by VS/GS in accessing constant, vertex or texture buffers in parallel with the data being streamed-out.

I hope the AMD and NVidia developer sites get some presentations from GDC soon...

Jawed

Arun

09-Mar-2007, 23:47

When SO is used on the back end of a data-amplifying render pass you might have 20-40 bytes in (per triangle) but 500-2000 bytes out (I'm unclear on the maximum amount of data SO can write per triangle).Yes, if the programmer is a complete and utter retard, he deserves to get bad performance. Is that what you meant? ;) :)

If SO is used to output 500-2000 bytes per triangle for something else than a GPGPU workload, that algorithm needs to be sent to the trashcan imo, and pronto! I agree it can be bandwidth intensive though, although since it also implies you're not doing stuff with the framebuffer at the same time, I would *hope* it's not too bad for real-world usages.

Jawed

09-Mar-2007, 23:51

Yes, if the programmer is a complete and utter retard, he deserves to get bad performance. Is that what you meant? ;) :)

If SO is used to output 500-2000 bytes per triangle for something else than a GPGPU workload, that algorithm needs to be sent to the trashcan imo, and pronto! I agree it can be bandwidth intensive though, although since it also implies you're not doing stuff with the framebuffer at the same time, I would *hope* it's not too bad for real-world usages.
So, what you're saying is that SO sucks on G80?

Jawed

IbaneZ

10-Mar-2007, 00:01

Fuad is wrong as usual. His site won't last long.

Well then Hector, please update us. :wink:

Razor1

10-Mar-2007, 00:27

So, what you're saying is that SO sucks on G80?

Jawed

Any GPU if doing that much out put from SO will get crushed in a real time environment is what Arun is hinting at.

Shtal

10-Mar-2007, 05:36

The GeForce 8800 Ultra may become an overclocked version of the GeForce 8800 GTX, which has been shipping for several month, however, considering the exceptionally low availability of the GeForce 7800 GTX 512, which was released in late 2005 and which featured an overclocked by 120MHz GeForce 7800 GTX chip, the new Ultra may not feature extreme clock-speed, but have other improvements compared to the model 8800 GTX. http://www.xbitlabs.com/news/video/display/20070305080809.html

I have 2 questions?

A. Probably G80-shrink using 80nm tech, plus running 650MHz GPU frequency including 25% more shaders.
B. How much pressure could it have on R600-XTX.

Russell

10-Mar-2007, 07:25

I have 2 questions?

A. Probably G80-shrink using 80nm tech, plus running 650MHz GPU frequency including 25% more shaders.
B. How much pressure could it have on R600-XTX.

A. No. Neither. That wouldn't be an 8800 Ultra. That'd be a 8900 GTX or something. I think it'd most likely be faster memory to try to keep pace with R600's 512-bit bus and perhaps moderately higher clock speeds (perhaps even a faster clock domain for the shaders).

B. Lots, regardless of the above (at least from the sounds of things).

Shtal

10-Mar-2007, 07:36

How much performance "percentage" do you expect to see out of Ultra version of G80?

Example: is it similar comparison (G70) 7800GTX vs. 7800GTX-512 ?

Russell

10-Mar-2007, 08:36

How much performance "percentage" do you expect to see out of Ultra version of G80?

Example: is it similar comparison (G70) 7800GTX vs. 7800GTX-512 ?

Personally? 5-10%. Probably closer to 10, which I think will compete nicely with R600.

Sound_Card

10-Mar-2007, 17:07

Personally? 5-10%. Probably closer to 10, which I think will compete nicely with R600.

How can you say that when we have no idea what the R600 is?

I seem to recall constant bad news for G80 like...
1) DX10 stapled on 7900
2) major power consumption
3) no unifide shader arc.

Then just mere day's befor G80 came out, it slapped evryones face.

Low behold, what's happening with R600?:razz:

I don't know what else other than that I highly doubt R600's 5/6 month "delay" was because of "silicon bugs".

Razor1

10-Mar-2007, 17:13

if it wasn't due to bugs or some sort of problems what did cause the 6 months in delays?

AlexV

10-Mar-2007, 17:26

The satisfaction of seeing heaps upon heaps of stupid questions and silly speculation being flung around:) In six months, those things tend to add up, and I`d bet that they`ve had a helluva lot of fun till now...that or there actually were some problems, which is not possible, coz ATi...erm, AMD, said there weren`t any, so my vote goes to for shits and giggles.

Kaotik

10-Mar-2007, 17:40

Of course it's a longshot, but is it out of the question that R600 is actually what was going to be R650 or R680 or whatever, the refresh anyway? It would at least explain the 5+ months delay

3dcgi

10-Mar-2007, 17:50

But it's worth noting that the output capabilities of the GS unit have been set "low" (4096 bytes of output data per input triangle, but 2048 bytes, I think, streamed out), in order not to push the hardware too much, either with internal post-transform cache or SO tiling/bandwidth usage...
I'm pretty sure you can streamout 4096 bytes as well.

trinibwoy

10-Mar-2007, 17:52

Of course it's a longshot, but is it out of the question that R600 is actually what was going to be R650 or R680 or whatever, the refresh anyway? It would at least explain the 5+ months delay

Given that every single time people have made that guess they were wrong, I'm gonna bet on it being wrong again. There has never been any fancy explanation for these things - it's always been very simple. The thing is just delayed, end of story.

trinibwoy

10-Mar-2007, 17:54

The satisfaction of seeing heaps upon heaps of stupid questions and silly speculation being flung around:)

Yeah last time I checked there was an entry for "satisfaction of seeing heaps upon heaps of stupid questions and silly speculation being flung around" right below gross profit on the income statement :roll:

AlexV

10-Mar-2007, 17:59

You ppl need to cheer up.

trinibwoy

10-Mar-2007, 18:01

Then just mere day's before G80 came out, it slapped evryones face.

Exactly, with all this stuff going on we still have no idea what R600 is like. Boards are up and running and info is still on lockdown. If R600 is as radical a departure from the norm as G80 was then it will be a fun launch...whenever it happens!

trinibwoy

10-Mar-2007, 18:04

You ppl need to cheer up.

An R600 review or two would do just that :)

Shtal

10-Mar-2007, 19:28

How can you say that when we have no idea what the R600 is?.

I think that's what people believe how R600 will scale against G80. :)

Russell

10-Mar-2007, 19:59

How can you say that when we have no idea what the R600 is?

I was asked what I expect. People have expectations without necessarily having the knowledge to back them up. That's more or less what I expect. I make no claim beyond that because, quite frankly, I couldn't back it up if I tried.

Sound_Card

10-Mar-2007, 20:07

if it wasn't due to bugs or some sort of problems what did cause the 6 months in delays?

No, I don't seem to recall ATi or AMD for that matter ever making first hand statements about R600's release. The Nov/Dec time frame was all the INQ part and say. Actully the offical statment regarding R600 release date was Cebit in late Q1. And the only offical "delay" would be now.

Russell

10-Mar-2007, 20:25

trinibwoy

10-Mar-2007, 20:55

No, I don't seem to recall ATi or AMD for that matter ever making first hand statements about R600's release. The Nov/Dec time frame was all the INQ part and say. Actully the offical statment regarding R600 release date was Cebit in late Q1. And the only offical "delay" would be now.

You know that's a weak statement. Companies never "officially" declare delays. Did you miss the recent conference where AMD pretty much said that the product is delayed but that they're confident that it will impress upon launch. Isn't it way too late to still be in denial about this whole thing?

Razor1

10-Mar-2007, 20:59

No, I don't seem to recall ATi or AMD for that matter ever making first hand statements about R600's release. The Nov/Dec time frame was all the INQ part and say. Actully the offical statment regarding R600 release date was Cebit in late Q1. And the only offical "delay" would be now.

Well if there were respins from A12 to A13 and possibly beyond there had to be some fixes :wink: , why would they waste money and time for another spin?

So ATi planned for a year and half without a new next gen high end product is what your saying, so they can getting dragged across the floor? A delay was definitly there if they announced it offically or not, I don't remember any official annoucement when the nv30 was delayed in the first Q either, once ya start going past that, people start asking questions and the something must be said.

Kaotik

11-Mar-2007, 03:26

have we actually seen chips with a12 or a13 or a-anything for that matter? or are they just rumors?

Shtal

11-Mar-2007, 07:25

Exactly, with all this stuff going on we still have no idea what R600 is like.

I made small poll; feel free to vote about what your expectations are....
Just for fun!!!! :) :) :) :) Thanks, http://snappoll.com/poll/177165.php

best regards,

Shtal....

Mintmaster

11-Mar-2007, 10:24

When SO is used on the back end of a data-amplifying render pass you might have 20-40 bytes in (per triangle) but 500-2000 bytes out (I'm unclear on the maximum amount of data SO can write per triangle).Jawed, first of all you missed my point entirely. How are you going to generate that much output data without consuming a lot of shader cycles? 2000 bytes output size is fine for G80's BW if you need 4000 scalar instructions to create it.

Second of all, fundamentally one is not changing the basics of 3D rendering when using SO. You still need some bizarre pixel shader that will use such huge amounts of iterator data. You still need to radically change the workload so that vertex bandwidth jumps like crazy from what we see today.

High BW from SO usage will only be found in corner cases and contrived examples. It is not a reason for a huge bump in BW.

Razor1

11-Mar-2007, 16:50

have we actually seen chips with a12 or a13 or a-anything for that matter? or are they just rumors?

a12 and a13 are pretty much confirmed

CarstenS

11-Mar-2007, 18:20

Maybe I'm just not used to seeing triple digit bandwidth numbers as yet but I have a hard time understanding why or how R600 would possibly use more than 115GB/s.
Given - i'm going to paint a very very worst-case picture here:

- Remember Xenos and how it's ROPs are off-die in the eDRAM-Module?
- Remember Rumors of earlier silicon of R600 having the most severe problems with its ROPs?

Now, is it very hard to imagine to have them use less than optimal ROPs and compensating for it with sheer bandwidth?

11-Mar-2007, 23:41

Jawed

11-Mar-2007, 23:45

Jawed, first of all you missed my point entirely.
My original point was about SO's sensitivity to bandwidth, not bandwidth consumption per se.

e.g. if SO uses uncached writes to memory (I'm not saying it does) then the performance would be horrible - but it would prolly be cheap to implement and it would be functionally complete.

SO would end up extremely sensitive to bandwidth/latency and well, prolly a dozen and one other things.

Jawed

INKster

11-Mar-2007, 23:53

Some more info leaked. Supposedly AMD set a new launch timeframe: 27 March - Early May. So that's a huge timeframe. Final launch driver is 8.361 (so Fuad (http://www.fudzilla.com/index.php?option=com_content&task=view&id=81&Itemid=1) was right about this one for once). RV630 should still be on track for a May launch, while RV610 will most probably be delayed. No reasons given for the delay. Some info was given about a possible counter from nV, but not sure how much credit I should give this piece of information (something about it being a 1GB card, source not sure if it's a G81 or a G80U which I didn't think would be possible to come in a 1GB version)

With a 384bit bus, yes, it's hard to conceive a 1GB card.

On the other hand, if there is in fact a "G81", then it's reasonable to assume that both the memory controller and the PCB could be tweaked to allow a 512bit bus, hence providing the possible 512MB and 1GB capacity versions of the future ultra high end cards.

Personally, i believe this is not the case, and any possible counter-strike from Nvidia would be a highly clocked derivative of the 90nm G80 with 1.5GB of fast GDDR4.

trinibwoy

12-Mar-2007, 19:24

I'm gonna put in a vote for higher clocked 80nm version with 768MB GDDR4 in June-July. 1.5GB is just too unnecessary and expensive for little benefit.

armchair_architect

13-Mar-2007, 02:42

if SO uses uncached writes to memory (I'm not saying it does) then the performance would be horrible

Why? Streaming output seems to be a poster child for not needing a cache: large coherent writes to sequentially increasing addresses. All you need is enough fifoing to coalesce the individual writes into full memory bursts.

icecold1983

13-Mar-2007, 03:06

hopefully r600 offers higher af, 32x would be great

Unknown Soldier

13-Mar-2007, 03:57

AMD mixes DX9 and DX10 GPUs in X2K series (http://www.theinquirer.net/default.aspx?article=38158)

At first, it seemed that some partners jumped the gun and prelaunched AMD's Mobility Radeon X2300 (allegedly an RV610 part) but, in the end, it turns out that AMD will mix its present DirectX 9 and future DirectX 10 parts in its X2K lineup.

The X2300 is actually just a rebranded X1300, and the probable reason for the change was a performance delta between 690G with X1250 graphics and the X1300 - X1250 is based on old X300 chip, while X1300 was newer RV510 part. AMD marketeers thought it would be a good idea to separate the high-performing part from the lower ones.

SugarCoat

13-Mar-2007, 04:47

hopefully r600 offers higher af, 32x would be great

disagree, if anything we need some improvement to mipmap levels.

Pete

13-Mar-2007, 06:11

1.5GB is just too unnecessary and expensive for little benefit.
The inevitable GX2 is gonna feel miiighty insulted when that reaches its sensitive ears. :)

_xxx_

13-Mar-2007, 10:37

Now, is it very hard to imagine to have them use less than optimal ROPs and compensating for it with sheer bandwidth?

Well yes, it's hard to imagine :)

trinibwoy

13-Mar-2007, 13:16

The inevitable GX2 is gonna feel miiighty insulted when that reaches its sensitive ears. :)

Heh, if there is a GX2 (which I doubt) it'll probably be two 512MB cards bolted together - not two full GTX's :)

Jawed

13-Mar-2007, 18:08

Why? Streaming output seems to be a poster child for not needing a cache: large coherent writes to sequentially increasing addresses. All you need is enough fifoing to coalesce the individual writes into full memory bursts.
I don't think they're necessarily large coherent writes:

up to 4 independent streams written in parallel
as little as 32 bits written per invocation, per stream
invocations potentially spread across all pipelines within the GPU, meaning that sets of invocations (batches) sent to each of the independently scheduled SIMD shader units can finish out of orderBut, frankly, the whole subject is so murky I dunno what to think.

"Cache" prolly isn't a good word, I admit, because it's always one-way data. Should have said buffered. I think the nature and scale of the write-buffering will make a difference to the performance of SO (that's always been my point, not exactly novel :smile: ). I don't know whether buffering within the memory controller would be enough to provide the maximum performance.

I wrote "uncached writes" as an allusion to the "scatter" support seen in X1k GPUs, which is known to perform slowly as it operates "uncached". Implying, to me, that all writes that don't go through the ROPs happen piecemeal. But I've not seen a description of how R5xx implements such writes.

Scatter is clearly more random than SO, but I'm not convinced SO is as coherent as it first appears.

Jawed

Arnold Beckenbauer

13-Mar-2007, 22:57

http://www.gecube.com/press-release-detail.php?id=65433
¡¯Exclusive GECUBE Product with the ATI World¡¦s most powerful Chipset
GECUBE will be unveiling at CeBIT its secret weapon (NDA) for the second quarter of 2007. Using the latest ATI core with support for the new generation DX10 technology, don¡¦t miss this chance to witness its jaw-dropping performance for yourself at CeBIT!

Pardon?

Bouncing Zabaglione Bros.

13-Mar-2007, 23:06

http://www.gecube.com/press-release-detail.php?id=65433

Pardon?

You won't be seeing it unless you're buying them in the thousands:

Apart from these products, GECUBE will also be unveiling its full series ultimate 3D graphics card using the latest chipsets from ATI with support for the new generation DX10 technology to its VIP business partners in its confidential (NDA) room. Don't miss out on this unique opportunity from GECUBE at CeBIT!

icecold1983

13-Mar-2007, 23:06

global 4x ssaa would put all that bandwidth to use, and would provide amazing iq

trinibwoy

13-Mar-2007, 23:27

global 4x ssaa would put all that bandwidth to use, and would provide amazing iq

Hmmmm, you might want the shading and texturing power to back that up too ;) Note 4xSSAA is like rendering the screen at 4x the resolution so everything in the core is taxed x 4 not just bandwidth...

icecold1983

14-Mar-2007, 00:00

stop raining on my parade with ur negativity, ur god damn negativity. im an idea man i thrive on enthusiasm

trinibwoy

14-Mar-2007, 00:06

lol sorry, i try to spend most of my time in the real world

/heaps rainbows and butterflies on icecold :p

icecold1983

14-Mar-2007, 00:14

if u rly think about it, its kind of odd that we have trouble rendering lifelike images in realtime when we can do such amazing things in other areas of technology.

Bouncing Zabaglione Bros.

14-Mar-2007, 00:21

if u rly think about it, its kind of odd that we have trouble rendering lifelike images in realtime when we can do such amazing things in other areas of technology.

We probably can if you wanted to pay tens of thousands for a graphics card. The technology isn't quite there yet if you want it in a (relatively) cheap card for your PC.

I'd be surprised if R600 will burn it's bandwidth on SSAA. I don't think AMD will turn away from "smarter" rather than "brute force" AA mechanisms. Developing the 512bit bus would have been too expensive if it was just for stepping backwards to a less efficient form of AA.