Huddy says "R600"

[trumphsiao]

With only 500M transistors?
They've got to be clocking it pretty fast, then, no?


some problem lies in ATI always want to do everything on their own without procuring any IPs

to speeding up whole progress .

[pjbliverpool]

Hard to beleive that a pair of these things in crossfire could actually have more bandwidth than Xenos's eDRAM all to a full 1GB of memory :shock: !

Incidentally, I don't expect R600 to come with only 64 Xenos style unified shaders, it isn't enough.

At 650Mhz it would only have as much total shader power as R580 has in its pixel shaders alone. Add in the vertex shaders and R600 would need to run at 740Mhz just to match R580's total shader power. That seems far too little for a next gen, DX10 part with a monster 512bit memory bus.

I expect 96 units at 600+ Mhz.

[Bouncing Zabaglione Bros.]

Hard to beleive that a pair of these things in crossfire could actually have more bandwidth than Xenos's eDRAM all to a full 1GB of memory :shock: !

Incidentally, I don't expect R600 to come with only 64 Xenos style unified shaders, it isn't enough.

At 650Mhz it would only have as much total shader power as R580 has in its pixel shaders alone. Add in the vertex shaders and R600 would need to run at 740Mhz just to match R580's total shader power. That seems far too little for a next gen, DX10 part with a monster 512bit memory bus.

I expect 96 units at 600+ Mhz.

I aggree. It might be plausable if this was a first attempt at a unified architechture, but after getting it right with Xenos, there's no way that R600 will only be on par with the R580 generation product. It's got to do a lot better than that.

[trinibwoy]

Well did they get it right with Xenos though? Is its performance impressive given clock speed and number of units?

[Geo]

At 650Mhz it would only have as much total shader power as R580 has in its pixel shaders alone. Add in the vertex shaders and R600 would need to run at 740Mhz just to match R580's total shader power.

Umm, what? Could you spell out the math a little more explicitly on how you got there? I'm not seeing any way a 650mhz 64 shader R600 doesn't smoke an R580 pretty thoroughly (tho would feel awfully imbalanced with 512-bit gddr4 unless we're missing something else).

[trinibwoy]

Umm, what? Could you spell out the math a little more explicitly on how you got there? I'm not seeing any way a 650mhz 64 shader R600 doesn't smoke an R580 pretty thoroughly (tho would feel awfully imbalanced with 512-bit gddr4 unless we're missing something else).

Play the rough theoretical FLOP game. 650*9*64 (374) for R600 vs 650*(48*8+8*10) (302). I think Xenos loses an ADD per shader compared to R580. So yeah it does look like by the numbers R600 should come out on top. Don't know about smoking though.

Edit: Corrected calculation based on Dave's feedback.

[pjbliverpool]

Umm, what? Could you spell out the math a little more explicitly on how you got there? I'm not seeing any way a 650mhz 64 shader R600 doesn't smoke an R580 pretty thoroughly (tho would feel awfully imbalanced with 512-bit gddr4 unless we're missing something else).

Sure. We know R580 packs 374.4 GFLOPS in the pixel shaders and another 52 in the vertex shaders for a total of 426.4 GFLOPS across all shaders.

So for R600 to use 64 Xenos style shaders would mean 64*9 FLOPS per cycle or 576 FLOPs per cycle.

576*0.65 = 374.4 GFLOPS

[Dave Baumann]

Xenos is Vector MAD + Scalar ADD (and special function). 216GFLOPs.

[pjbliverpool]

Play the rough theoretical FLOP game. 650*10*64 (416) for Xenos vs 650*(48*8+8*10) (302). I think Xenos loses an ADD per shader compared to R580. So yeah it does look like by the numbers R600 should come out on top. Don't know about smoking though.

The R580 ALU's make up 12 FLOPS including the ADD while Xenos's Scalar is not MADD capable and thus only accounts for 1 FLOP (for a total of 9) - I realise it may be capable of more under certain circumstances but 9 is probably the best figure to use for a Xenos ALU.

So its more like 650*9*64 (374.4) vs 650*(48*12+8*10) (426.4).

[Geo]

Doh, I forgot to credit for the less capable ALU in R580 ps. Tho I was crediting 20% for efficiency pickups in R600, re what Huddy reported way back up near the top.

[trinibwoy]

The R580 ALU's make up 12 FLOPS including the ADD while Xenos's Scalar is not MADD capable and thus only accounts for 1 FLOP (for a total of 9) - I realise it may be capable of more under certain circumstances but 9 is probably the best figure to use for a Xenos ALU.

Cool, thought the ADD was MADD capable as well. How'd you get 12 flops per R580 shader though?

Doh, I forgot to credit for the less capable ALU in R580 ps. Tho I was crediting 20% for efficiency pickups in R600, re what Huddy reported way back up near the top.

Don't you mean more capable :wink:

[Geo]

http://www.beyond3d.com/reviews/ati/r580/index.php?p=03

About 1/2 way down. ALU1 is 4 (ADD), ALU2 is 8 (MUL+ADD)

No, by "less capable" I meant I forgot to include ALU1 in my calcs. :sad:

[pjbliverpool]

Doh, I forgot to credit for the less capable ALU in R580 ps. Tho I was crediting 20% for efficiency pickups in R600, re what Huddy reported way back up near the top.

Is it really more efficient though?

Im no expert so I would love to here the devs views on this but I think its at least possible that while overall ALU usage is up (obviously) in a unified system, the individual ALU's are less efficient due to the lack of specialization. Wouldn't these two factors counteract each other somewhat?

[trinibwoy]

http://www.beyond3d.com/reviews/ati/r580/index.php?p=03

About 1/2 way down. ALU1 is 4 (ADD), ALU2 is 8 (MUL+ADD)

No, by "less capable" I meant I forgot to include ALU1 in my calcs. :sad:

Oh my bad. Thanks for the link.

[Geo]

Is it really more efficient though?

Im no expert so I would love to here the devs views on this but I think its at least possible that while overall ALU usage is up (obviously) in a unified system, the individual ALU's are less efficient due to the lack of specialization. Wouldn't these two factors counteract each other somewhat?

They might, but not completely. Or they wouldn't have done a *second* USC would they? I mean, maybe they could be wrong on theory vs practice on the first one, but they'd have to be pretty stupid to do the second one wouldn't they?

[trinibwoy]

Why don't we ever have numbers to go with the supposition that unified is more efficient :sad: Console devs speak up !!

[pjbliverpool]

They might, but not completely. Or they wouldn't have done a *second* USC would they? I mean, maybe they could be wrong on theory vs practice on the first one, but they'd have to be pretty stupid to do the second one wouldn't they?

True, but raw power isn't the only advantage of the USC.. im thinking we have to go that way regardless because of the developer flexibility it affords.

In the end I assume it will simplify GPU design aswell.

[Geo]

How would you measure that other than how Huddy reported it up at the top, by locking in the ratio vs letting it float?

I guess it's kind of interesting that tho you have 50% more PS and 200% more VS peak capacity, the actual performance improvement was 20-25%. Of course was there BW limitations at play?

The other potentially interesting question is R580 has two asymmetric ALU's per ps. Possibly that means there is a "wastage" there by making all ALUs identical in a USC to offset to some degree the dynamic allocation advantages. The argument against that tho is that NV moved towards greater symmetry with their PS ALU's going from NV40 to G70.

[trinibwoy]

How would you measure that other than how Huddy reported it up at the top, by locking in the ratio vs letting it float?

Well locking in a ratio isn't a valid comparison since the architecture wasn't designed to be locked in in the first place. What I meant was that devs should have a good feel for Xenos and RSX by now. That's 64 unified vs 56 specialized (48+8). I think somebody should be able to make a comparison by now. Huddy isn't exactly an objective source!

[Geo]

Well locking in a ratio isn't a valid comparison since the architecture wasn't designed to be locked in in the first place. What I meant was that devs should have a good feel for Xenos and RSX by now. That's 64 unified vs 56 specialized (48+8). I think somebody should be able to make a comparison by now. Huddy isn't exactly an objective source!

Well, no he isn't. But then again nobody made them make that switch. They had an architecture that wasn't unified, and chose to move to one that was. No "Huddy vs NV" type of analysis is necessary. I'm talking about "ATI vs ATI" here.

As has been noted many times, the programming interface would be transparent to devs either way. The only real advantage to the devs would be more predictably scalable performance, which probably isn't enough to make that move all by itself if it's not bringing you performance advantages outweighed by its drawbacks. Certainly not to do it a second time.

There surely are console devs supporting both XB360 and PS3, but you'd really have to wonder if the other variables outside the gpus make that all that useful a comparison to try to make for *just* the gpu piece.

[trinibwoy]

Well, no he isn't. But then again nobody made them make that switch. They had an architecture that wasn't unified, and chose to move to one that was. No "Huddy vs NV" type of analysis is necessary. I'm talking about "ATI vs ATI" here.

That - "we should only compare architectures from the same company" - argument doesn't fly. We know the number of shaders, the general capabilities of each and the clockspeed. That is more information than most people use to make a judgment on superiority. If we can get feedback on relative performance of Xenos vs RSX then we will be able to extrapolate that to the PC space. There are assumptions there as well but I'd take that over a comparison that involves crippling the architecture.

As has been noted many times, the programming interface would be transparent to devs either way. The only real advantage to the devs would be more predictably scalable performance, which probably isn't enough to make that move all by itself if it's not bringing you performance advantages outweighed by its drawbacks. Certainly not to do it a second time.

Well the argument that ATI wouldnt build R600 if Xenos didn't work is one thing (the assumption being that they could do anything about it anyway). I'm just talking about simple, run-of-the-mill performance comparisons.

There surely are console devs supporting both XB360 and PS3, but you'd really have to wonder if the other variables outside the gpus make that all that useful a comparison to try to make for *just* the gpu piece.

Well given feedback from developers here I'm certain that these guys do a lot of focused testing and don't just build a game and then start worrying about performance. Of course these guys know how fast individual hardware components are on specific tasks.

[Geo]

That - "we should only compare architectures from the same company" - argument doesn't fly.

Nah, that's not what I meant there. Just addressing Huddy's "bias" against (presumably) R5xx in favor of R6xx.

Now, towards the end that's what I meant, complicated by Cell vs Xenon, different OS/API, etc. :razz:

[trinibwoy]

Nah, that's not what I meant there. Just addressing Huddy's "bias" against (presumably) R5xx in favor of R6xx.

Now, towards the end that's what I meant, complicated by Cell vs Xenon, different OS/API, etc. :razz:

Bah, guess I'll just have to wait till February :)

[pjbliverpool]

Well, no he isn't. But then again nobody made them make that switch. They had an architecture that wasn't unified, and chose to move to one that was. No "Huddy vs NV" type of analysis is necessary. I'm talking about "ATI vs ATI" here.

Im not sure how true it is, but wasn't there some kind of internal conflict in ATI over whether to make R580 unified or not? There may be other factors at play but that suggests to me that the performance advantage of unified over seperate can't have been huge (or even existent) with that transistor budget given that they settled on seperate.

I could be way off base though.... Im sure at least a few people here know the answer :wink:

[hoom]

Heck, by my understanding (no inside info) there was internal conflict over whether R4xx would be unified...

[Geeforcer]

Inquirer rumor about a next-gen ATI card? Then it must be true (http://www.theinquirer.net/default.aspx?article=22202).

[Rangers]

Well locking in a ratio isn't a valid comparison since the architecture wasn't designed to be locked in in the first place. What I meant was that devs should have a good feel for Xenos and RSX by now. That's 64 unified vs 56 specialized (48+8). I think somebody should be able to make a comparison by now. Huddy isn't exactly an objective source!

Xenos is 48 unified ALU's, not 64.

So in that sense, RSX supposedly actually has more raw power (although there are so many caveats, it's almost impossible to make any comparison).

If you could draw one conclusion, it would probably be that unified has NOT gained Xenos a vast amount of efficiency, it doesn't blow away RSX by the (current?) games, despite, similar die sizes and so forth.

[vertex_shader]

Inquirer rumor about a next-gen ATI card? Then it must be true (http://www.theinquirer.net/default.aspx?article=22202).

Its not Inqurier rumor, its come from asia http://www.google.com/translate?u=http%3A%2F%2Fwww.hardspell.com%2Fnews% 2Fshowcont.asp%3Fnews_id%3D30133&langpair=zh-CN%7Cen&hl=en&ie=UTF8

[Ailuros]

Light travels faster than sound; that's why some trash tabloid authors appear bright until they open their traps (sorry couldn't resist...).

[trinibwoy]

Xenos is 48 unified ALU's, not 64.

So in that sense, RSX supposedly actually has more raw power (although there are so many caveats, it's almost impossible to make any comparison).

If you could draw one conclusion, it would probably be that unified has NOT gained Xenos a vast amount of efficiency, it doesn't blow away RSX by the (current?) games, despite, similar die sizes and so forth.

Ah true. But you got the gist of what I was saying.

[Geo]

Inquirer rumor about a next-gen ATI card? Then it must be true (http://www.theinquirer.net/default.aspx?article=22202).

Oh, that's a piker effort. VR-Zone had actual yield numbers for 32/24 pipe R520! :grin:

[Ailuros]

Ah true. But you got the gist of what I was saying.

I've seen developers note on quite a few occasions that it's very easy to code for XBox360; I'd figure though that it's not directed exclusively to the graphics chip of the console.

[Ailuros]

Oh, that's a piker effort. VR-Zone had actual yield numbers for 32/24 pipe R520! :grin:

I'd love the conversation to concentrate on the possibility of a 512bit bus a tad more, even if it's purely from the theoretical technical POV. Take my layman's perspective a given, but my primary concern with such wide bus is how they'd handle read efficiency, which if not handled properly might even reduce the overall available bandwidth. Granted 384bits is not going to be easy either for G80, but the hypothetical 512bits are another notch higher.

[Geo]

Well, it seems to me that 384/320 for NV and Sireric's comments on dedicating clients might suggest the answer there. That was the theory behind the crossbar in the first place wasn't it? So I don't see these companies giving up that kind of granularity, but rather adding a couple more units to it.

[trumphsiao]

Well, it seems to me that 384/320 for NV and Sireric's comments on dedicating clients might suggest the answer there. That was the theory behind the crossbar in the first place wasn't it? So I don't see these companies giving up that kind of granularity, but rather adding a couple more units to it.



Maxim consumption of R600 based on exteme clock speed is less than 200W~250W

But hope this time ATI will prevail on Nvidia from Performance/mm^2 stance.

[Shtal]

I' am going to ask a question; which is still bothers me.

What is the main reason that R600 is delay for 3 months after the G80 will be released.

A.) Is it because ATI is behind on R600 project and Nvidia is ahead with G80.

B.) Is it because R600 is going to be more advanced then G80, so ATI needs more time.

C.) Or ATI has hard-time solving problems with R600 just like they did with R520 and it was delay many times.
1.) GPU frequency MHz
2.) Manufacture process
3.) Low yields

If the answer is C.
Then Nvidia finally did it for sure, they finely became leading technology again.

[trinibwoy]

Somehow I doubt you're going to get an answer anytime soon. So unfortunately you're going to be bothered for a bit longer :)

[Geo]

That kind of conclusion is fairly facile anyway. What would you want to say about G71 being pretty much a straight die shrink of G70? 1). NV's tech was so uber they could stand pat with a process bump. 2). They mailed it in to focus on G80 instead after G70 launch.

I wouldn't be comfortable saying either, but YMMV.

It does sound pretty lucky for ATI that NV didn't hit their original dates for G80 tho, or they'd really be screwed.

Edit: It would be interesting to know when X1950 showed up as definitely on the schedule, as that seems to be a key inflection point.

[Geeforcer]

As geo pointed out, within last two years ATI went

R420 -> R520/580 -> R600

While we don't know much about R600, I'll venture a guess that it's a fairly major revision ;).

Nvidia on the other hand has been milking NV40 for last two years. Nothing wrong with that, (god knows ATI did the same with R300) but it does mean they had more time to get G80 ready, since it will be their first major core revision in the last two years.

[SugarCoat]

I don't know that I'd ever heard a specific date range from ATI.

Nobody at a conference call said they were on schedule for a Q4 launch around the beginning of this year? Its so vague anyway though i guess it doesnt matter, especially when the product, a very complex one at that, isnt finalized yet. Is there a delay? Is it an expected delay do to a scheduled change? Or did they crap out on yields for target clocks, blah blah... Actually quite useless speculating.


As geo pointed out, within last two years ATI went

R420 -> R520/580 -> R600

While we don't know much about R600, I'll venture a guess that it's a fairly major revision ;).

Nvidia on the other hand has been milking NV40 for last two years. Nothing wrong with that, (god knows ATI did the same with R300) but it does mean they had more time to get G80 ready, since it will be their first major core revision in the last two years.


Well current products and past products arent exactly a map for what the future holds in store. I think the R300 and even the failure of the FX series have proven that quite well. We dont exactly know how much lineage a core contains in comparison to its predecessor, so coming to the conclusion that Nvidia has had a longer time isnt exactly as simple as that. Even Xenos R&D started years ago, before the NV40 launched.

ATI has technically been milking it longer though without a truly major architecture revision. R580 for sure was interesting but things like AA and even AF (barring the addition of HQ of course) are essentially the same things done by the R300, so R300 -> Present > NV40 -> NV50;)

[INKster]

As geo pointed out, within last two years ATI went

R420 -> R520/580 -> R600

While we don't know much about R600, I'll venture a guess that it's a fairly major revision ;).

Nvidia on the other hand has been milking NV40 for last two years. Nothing wrong with that, (god knows ATI did the same with R300) but it does mean they had more time to get G80 ready, since it will be their first major core revision in the last two years.

I wouldn't be so sure about that...

ATI has had basically a new architecture every 2/2.5 years.
-R200->R250->R280.
-R300->R350->R360->R420->R480.
-R520->R580.

With the exception of GeforceFX/NV3x (because of it's poor performance and delays, it lasted little more than a year) so has NV:
-NV10->NV15.
-NV20->NV25/NV28.
-NV40/NV45->G70->G71.


So, if R600 is really such a leap forward compared to the previous generation (on the architecture level, at least), it will be a tight fit on the schedule, just over a year after R520's debut.
With the exception of unified shaders and DX10, i believe more that they'll just keep the investment in performance using the current, proven, memory controller, add ALU's and/or clockspeed and that will be it until the DX10.1 refresh (R680 ?) and then the true new design will be R700.

Of course, either architecture may even turn out to be totally "new" (sort of speak), and it still will be up to driver support and scalability of each into the refresh parts 6 to 8 months later, and beyond.

[Dave Baumann]

So going from a non-unified architecture to a unified on doesn't constitute a major architecture change?

[INKster]

So going from a non-unified architecture to a unified on doesn't constitute a major architecture change?

Well, since G80 was being perceived as a minor change towards speed and DX10 (non-unified, 48 pipes, etc), and then took a quick turn to being unified and scalar (or so the rumour goes), i'm not ruling anything out yet.

But, there has been one unified chip on the market for a year, right ?
What's so different about R600 ?

Some nice tasty bone about it should be in order, at least to stir up a bit the discussion between G80 and R600. :wink:

[Shtal]

I was reading alot of comments (sort of) about R600.

What about name; since it is no longer ATI but AMD now.

It might be like this:

[Top Perfomer]
AMD Radeon X2800XT+ (Since AMD loves plus+ at the end)

[Performer]
AMD Radeon X2800XL+

Also it might be Green+Red box package Welcome new comer AMD video product.

[Arty]

(Since AMD loves plus+ at the end)
AMD adopted the + because the consumers thought Athlon processors were inferior based on just the clock speed. It was an ordinal comparison with Pentium processors.

Sorry I dont see that or the green+read package happening.

[Shtal]

Sorry I dont see that or the green+red package happening.

So what is going to be; your name instead:)

[trumphsiao]

I' am going to ask a question; which is still bothers me.

What is the main reason that R600 is delay for 3 months after the G80 will be released.

A.) Is it because ATI is behind on R600 project and Nvidia is ahead with G80.

B.) Is it because R600 is going to be more advanced then G80, so ATI needs more time.

C.) Or ATI has hard-time solving problems with R600 just like they did with R520 and it was delay many times.
1.) GPU frequency MHz
2.) Manufacture process
3.) Low yields

If the answer is C.
Then Nvidia finally did it for sure, they finely became leading technology again.


RV570/RV560 is overclockable unfriendly.

If R600 want to be as fast as 8800GTX. ATI need to launch 600MHz~750MHz R600 .

[trumphsiao]

That kind of conclusion is fairly facile anyway. What would you want to say about G71 being pretty much a straight die shrink of G70? 1). NV's tech was so uber they could stand pat with a process bump. 2). They mailed it in to focus on G80 instead after G70 launch.

I wouldn't be comfortable saying either, but YMMV.

It does sound pretty lucky for ATI that NV didn't hit their original dates for G80 tho, or they'd really be screwed.

Edit: It would be interesting to know when X1950 showed up as definitely on the schedule, as that seems to be a key inflection point.


Nvidia 80nm is overclockable compared to ATI 80nm medicore clock increase.

[Shtal]

We all are not sure how R600 will turn out; so we cannot come to conclusion what frequency R600 needs to run in order to match or outperform G80.

[trumphsiao]

We all are not sure how R600 will turn out; so we cannot come to conclusion what frequency R600 needs to run in order to match or outperform G80.



Actually both of them are equally performance on same clock speed. but the problem is if ATI launch R600 by 2007 Q1, Nvidia would be rehashing version of Higher-Clock G80(GDDR4 ) in 3 months by then.

[Arun]

Actually both of them are equally performance on same clock speed.That statement and such a blatant simplification is kind of an insult to mostly everything Beyond3D stands for, imo... I'm honestly tired of this kind of overconfidence again, again, and again. I know it's not really intentional, but that's not really a reason to continue doing so either.


Uttar

[trumphsiao]

That statement and such a blatant simplification is kind of an insult to mostly everything Beyond3D stands for, imo... I'm honestly tired of this kind of overconfidence again, again, and again. I know it's not really intentional, but that's not really a reason to continue doing so either.


Uttar


ok. I discount Higher Clock domain Unit on G80 .(200W~250W)R600 is faster than G80 (150W)

.:grin: sorry for my malfeasance.

[Shtal]

Actually both of them are equally performance on same clock speed. but the problem is if ATI launch R600 by 2007 Q1, Nvidia would be rehashing version of Higher-Clock G80(GDDR4 ) in 3 months by then.

Before Nvidia do refresh product in 3 months; they have to find out about R600 first.

A. Is it going to have 512-bit memory internal + external at the same time
B. Is it going to have over 2000MHz+ frequency GDDR4 memory
C. Is it true that it will have true 64 physical pipelines
D. Or it will catch-up to Nvidia specs first; barely match nvidia...

[Arun]

ok. I discount Higher Clock domain Unit on G80 .(200W~250W)R600 is faster than G80 (150W)

.:grin: sorry for my malfeasance.That's not the point. You're talking like you know everything about these architectures and had the possibility to benchmark both of them in person on similarly specced machines. But obviously, you haven't. Even with the full specifications, you have NO idea if the two chips would have identical performance per clock. You could easily claim the performance of R420 and NV40 were similar per clock if you "simplified" a few things, but guess what, they were NOT! So what you're saying is downright stupid.

Also, you keep talking like everything you say is fact. Yet the "facts" change every 3 days in your posts. I'm not saying part of your information isn't correct, and I'm sure some of it is. That's why it's still nice to see you posting what you think from time to time. But as it is, you are merely adding noise to the forum with short posts full of unreliable content, the whole thing added with a huge amount of confidence in information that you shouldn't be that certain about, even if your source was Dave Orton or Jen Hsun Huang themselves! There is no such thing as a fully reliable source, and it's about time you learn that. This is getting extremely annoying.

I'd like to see you contributing to these forums of course! But with a tad less... extreme posting habits. That is, it'd be nice if you contributed more than a ton of minuscule posts that are only right 30-40% of the time, maximum. While that's not a bad percentage, the fact you write it in your quite awful english (it's not your fault, I know... It just isn't very fun to read and understand though) and with that confidence is ridiculous. Even more so considering the amount of posts you make in a week... This is all my opinion, but I think others will most likely agree. As I said, we appreciate some of your contributions, but often your posting style and opinions can be VERY annoying for the rest of us! :)


Uttar

[Acert93]

Before Nvidia do refresh product in 3 months; they have to find out about R600 first.

A. Is it going to have 512-bit memory internal + external at the same time
B. Is it going to have over 2000MHz+ frequency GDDR4 memory
C. Is it true that it will have true 64 physical pipelines
D. Or it will catch-up to Nvidia specs first; barely match nvidia...

What I want to know is if R600 has any surprises up the ol' sleeve like G80 may, specifically divergently clocked domains. ATI seemed to make a big deal out of ******** a year or two ago and some comment about how it wasy extremely vital to ATI.

I would guess that we also don't know everything about R600. Sure, ATI will leverage Xenos' unified shaders and the memory controller in R5x0 (and both should be quite well tweaked and optimized), but those will be 15+ months old by the time R600 launches. Obviously a lot of time has to go into meeting the D3D10 spec and Xenos was a large side project, but I am expecting some surprises from ATI.

[Geo]

Oh damn, I wanted to see if we'd get to #1000 before a Fast14 mention! :wink:

[Acert93]

Oh damn, I wanted to see if we'd get to #1000 before a Fast14 mention! :wink:

Great, and now you go mentioning it out of the blue!

/me whistles walking away....

Seriously though, with how the industry goes in regards to technology in general, what is available to NV is available to ATI and vice versa. It is all about choice. Seeing as G80 may have ALUs clocked much higher than the rest of the chip, it does raise the question about unsaid technology agreement ;) The big selling points (http://http://www.extremetech.com/article2/0,3973,1518127,00.asp), made to the public, were "According to Intrinsity, circuit designers using Fast14 can be more productive, allowing them to design processors that can run at high clock rates" and "Intrinsity's Fast14 technology could enable ATI to design GPU logic that clocks up to four times faster than today's clock speeds, according to a company spokesman for Intrinsity. Current high end GPUs clock in the 400-600MHz range, so a chip using Fast14 technology could deliver GPUs that run in the 1.6 GHz to 2.4 GHz range. Fast14 doesn't help with memory clocks, including onboard cache memory, so only the logic portion of the chip would run at the high clock rates".

PR bull aside (it seems in reality the deal was about delivering more performance per silicon dollar), the fact we are seeing a G80 leaks indicating nearly 1.5GHz shader ALUs raises an eyebrow. In many ways G80 is sounding a lot like what many were expecting from R600: Unified Shaders, a bit of bandwidth, and ALUs with high clock speeds. Maybe someone switched up the PR slides :razz: If you had handed me the above info and asked me what GPU it was I would have said R600... Anyhow, I wouldn't be surprised if R600 was clocked very fast, and if G80 is what we think it is it better be!

[rwolf]

What I want to know is if R600 has any surprises up the ol' sleeve like G80 may, specifically divergently clocked domains. ATI seemed to make a big deal out of ******** a year or two ago and some comment about how it wasy extremely vital to ATI.

I would guess that we also don't know everything about R600. Sure, ATI will leverage Xenos' unified shaders and the memory controller in R5x0 (and both should be quite well tweaked and optimized), but those will be 15+ months old by the time R600 launches. Obviously a lot of time has to go into meeting the D3D10 spec and Xenos was a large side project, but I am expecting some surprises from ATI.

That's what I want to know! Intrinistiy was talking 2GHz and 2.5GHz for the 130nm process. There was a brief period where they had 4GHz listed for 65nm then they replaced it with greater than 3GHz without mentioning the process.

[Jawed]

RV570/RV560 is overclockable unfriendly.

If R600 want to be as fast as 8800GTX. ATI need to launch 600MHz~750MHz R600 .
Why's the speed of mainstream parts, RV5xx, relevant to R600?

That "V" in the name means they're built to a cost. The board itself will be built to a cost as well, which ultimately means there's no reason for ATI to build them for overclockability. Particularly as this far into the lifetime of R5xx the positioning of these SKUs is well understood. They know what they're aiming for, there's no need to aim higher.

Each process node has speed grades. It's reasonable to assume that RV5xx is using the GT grade of 80nm, while R600 will be using the high-speed grade of 80nm:

http://www.fabtech.org/content/view/1170

TSMC's high-performance GT process as used by the graphics chip companies is the first to enter production. In the following months TSMC plans to offer its high-speed HS process and low power LP processes at the 80nm node by March 2006. A special GC process, which provides both low active and standby power advantage, will become available in the third quarter of 2006, according to the company.
We should then see (some of?) the performance/mainstream/value R6xx variants being introduced on 65nm, leaving the enthusiast level R600 waiting for refresh, ~9 months later, before it hits 65nm...

http://www.beyond3d.com/forum/showthread.php?t=29612

Dave then went on to say that none of the R6xx generation is likely to be 90nm based, instead split between 80nm and 65nm processes. This suggests that ATI may be adopting the process choices they did with the R3xx and R4xx generation, by introducing the new architecture first at the high end on a known process, and moving the derivative, lower end parts to a newer, smaller process.
What's the timing on 65nm? How much of the timing of R600 is tied into the timing of a "big bang" R6xx range, with a mix of 80nm HS and 65nm GT parts releasing over the period of one quarter, say?

Jawed

[Arun]

Thread locked, as it became a tad messy (not more so than the G80 one, but we'll be fixing that soon too) - please use the new one (http://www.beyond3d.com/forum/showthread.php?t=34676) for all further R600/R6xx discussion!

Uttar