NVIDIA Fermi: Architecture discussion

[green.pixel]

[off]I think you confusing terms, XRay engine of STALKER is not so bad

I'm not saying it's bad at all. It's far from it actually. The perfomance is, to some extent, imo.
http://www.guru3d.com/article/stalker-clear-sky-graphics-card--vga-performance-roundup/4


It's like with Crysis and huge ammount of DP calls being made. I just think that both of these series could see some perfomance optimization. <shrugs>

 

[GZ007]

No, nVidia's advantage in tessellation comes from the higher geometry performance and the four triangles/clock. I think it's possible that we will see no performance hit with tessellation in stalker and a low one in the unigine benchmark.

I dont see your point here. From the dx10 285GTX scores from the hardocp page u can asume that GF100 will have much higher fps without tesselation in the heaven benchmark than any other card.
In the heaven benchmark slide they didnt showed the fps without tesselation for some reason
And that should be the 512 core card.

 

[Razor1]

there were no 512 sp cards at CES :smile:

 

[Tamlin]

there were no 512 sp cards at CES :smile:

Why?

 

[Squilliam]

there were no 512 sp cards at CES :smile:

How many of them so called 'Cootie Cores' did them kards have?

 

[Ailuros]

I dont see your point here. From the dx10 285GTX scores from the hardocp page u can asume that GF100 will have much higher fps without tesselation in the heaven benchmark than any other card.
In the heaven benchmark slide they didnt showed the fps without tesselation for some reason
And that should be the 512 core card.

You mean the cutscene they used out of the Unigine heaven demo obviously. It's a tesselation limited case afaik and in such a case I don't think you'd use a fillrate limited case f.e. to test/show the theoretical strengths of a solution for tesselation.

I think most of us can understand the meaning of synthetic measurements and cutscene or not I'd personally prefer the difference the Unigine demo seems to show for tesselation than NV's own tesselation oriented techdemos.

IMO that infamous cutscene result has received far more attention than it actually deserves. As long as people understand what PR/marketing aims for, cherry picked best case scenarios are expectable and I haven't seen anywhere in AMD's PR/marketing presentations any worst case scenarios either.

there were no 512 sp cards at CES :smile:

...errr and that's a good sign for what exactly? If that should be true I can draw more negative than positive conclusions out of that one if true.

 

[GZ007]

there were no 512 sp cards at CES :smile:

The heaven benchmark graph surely wasnt made at CES but inhouse on the fastest card they had with 512 sp and god know what frequency
Wouldnt you show your fastest card against competition if u want undecided buyers to wait and buy your card

 

[GZ007]

IMO that infamous cutscene result has received far more attention than it actually deserves. As long as people understand what PR/marketing aims for, cherry picked best case scenarios are expectable and I haven't seen anywhere in AMD's PR/marketing presentations any worst case scenarios either.

I was just saying that if GF100 has realy small performance drop with tesselation wouldnt it look better on graph to show both FPS with tesselation on and off
Cypress wouldnt look too smart there if GF100 would have something like 10% performance drop and the red card near 50%. I think this oportunity wouldnt be left out from nvidia PR/marketing if it was true

 

[Ailuros]

I was just saying that if GF100 has realy small performance drop with tesselation wouldnt it look better on graph to show both FPS with tesselation on and off
Cypress wouldnt look too smart there if GF100 would have something like 10% performance drop and the red card near 50%. I think this oportunity wouldnt be left out from nvidia PR/marketing if it was true

Essentially you could spin countless of scenarios based on each PR/marketing oriented illustration. My personal judgement on any new architecture/GPU out there falls only after I can read/evaluate a healthy number of independent measurements in an as large as possible variety of games.

Under that light try to make a list of games during 2010 for games with a DX11 path and the amount of tesselation that's going to be used; considering GF100 is aimed for a March launch there won't be much to show anyway. And that's exactly the reason why I feel that the results for the Unigine Heaven demo have been overanalyzed.

Since I as a user can hardly make much use of any tesselation hw pipeline during the year I'll leave it to developers to decide which implementation suits their needs better or ends up being more flexible or efficient and it won't in the least influence my buying decisions.

It's part of a marketing whitepaper/campaign and I can't see one good reason why I should be sold by it and that irrelevant if it would be AMD or NVIDIA.

***edit: if you really want to protest against carefully selected case scenarios here's a far better case: http://www.hardocp.com/image.html?image=MTI2MzYwODIxNHh4VHN0ekRuc2RfMV8xN19sLmdpZg==

 

[FrameBuffer]

Essentially you could spin countless of scenarios based on each PR/marketing oriented illustration. My personal judgement on any new architecture/GPU out there falls only after I can read/evaluate a healthy number of independent measurements in an as large as possible variety of games.

Under that light try to make a list of games during 2010 for games with a DX11 path and the amount of tesselation that's going to be used; considering GF100 is aimed for a March launch there won't be much to show anyway. And that's exactly the reason why I feel that the results for the Unigine Heaven demo have been overanalyzed.

Since I as a user can hardly make much use of any tesselation hw pipeline during the year I'll leave it to developers to decide which implementation suits their needs better or ends up being more flexible or efficient and it won't in the least influence my buying decisions.

It's part of a marketing whitepaper/campaign and I can't see one good reason why I should be sold by it and that irrelevant if it would be AMD or NVIDIA.

***edit: if you really want to protest against carefully selected case scenarios here's a far better case: http://www.hardocp.com/image.html?image=MTI2MzYwODIxNHh4VHN0ekRuc2RfMV8xN19sLmdpZg==

Of course noticably absent from that graph @ HardOCP is the mention that CUDA/PhyX was involved in addition to any tesselation... then again NV did say that their $150 card was capable of beating ATI's best and brightest (in CUDA/PhyX)..

 

[Ailuros]

Of course noticably absent from that graph @ HardOCP is the mention that CUDA/PhyX was involved in addition to any tesselation... then again NV did say that their $150 card was capable of beating ATI's best and brightest (in CUDA/PhyX)..

Who cares? The first 3 results are from "directed tests", two of them in house techdemos and the last "game state bucket". Any of those are clearly coded from scratch to show the maximum twisted benefit of their own hw and at the same time give a worst case for the competing solution to make it look as bad as possible.

Years ago PowerVR released one of the cutest benchmarks I've seen to date named Fablemark; now try to convince me that it wasn't a "directed techdemo" also and for a good reason extremely z/stencil fillrate limited. If an IHV can't make through their techdemos/benchmarks look any competing solution to look bad then what's the point of even bother coding it?

PR/marketing will always attempt to "sell" with hand selected best case scenarios and wild exaggerations; the only question here is who is "buying" exactly according to those.

 

[mczak]

I just looked it up, it is even said in the Evergreen ISA docs that FMA works for double precision parts only.

Yes, indeed (page 1-96). It is the only instruction which says that btw (not even the _64 instructions do, I guess it's implied there). Still a bit strange considering the supposedly minimal die area cost. Though I guess it'll also save a bit of power. And we don't really know what lower end fermi parts will offer on that front, while it certainly is possible to just scale it down they might be different, after all.

 

[trinibwoy]

I was just saying that if GF100 has really small performance drop with tesselation wouldnt it look better on graph to show both FPS with tesselation on and off

Nope, because that would only highlight its lack of an advantage in the "off" scenario. Silly thing to do when promoting your product

 

[GZ007]

Nope, because that would only highlight its lack of an advantage in the "off" scenario. Silly thing to do when promoting your product

It would also suck if GF100 would be just few fps faster without tesselation than cypress with all those improvements. So its more like a double edged sword.
Iam still convinced that GF100 will have similar scaling than cypress in heaven benchmark and the problem is in the engine. GF100 has 48 ROP-s and much better inter chip bandwith (and also 200+ GB offchip bandwith). It wouldnt make any sense not to have masive fps advantage without tesselation.

 

[mczak]

nvidia should have shown vantage perlin noise scores. If a GF100 beats a HD5850, I'd be impressed!

 

[itaru]

The handling of VS-HS-TS-DS-GS is speeded up by an effect of cache in GF100.

 

[Razor1]

The heaven benchmark graph surely wasnt made at CES but inhouse on the fastest card they had with 512 sp and god know what frequency
Wouldnt you show your fastest card against competition if u want undecided buyers to wait and buy your card

no it was not no 512 sp cards were sent prior to some time in the past month.

 

[Razor1]

...errr and that's a good sign for what exactly? If that should be true I can draw more negative than positive conclusions out of that one if true.

JPR's latest reports, stated that because of AMD's supply constraint due to TSMC issues, they lost desktop discrete marketshare to nV, which is quite interesting, how bad were the yields of AMD's 40nm chips, and imagine how a chip that is ~35% larger would be affected.

 

[Kaotik]

JPR's latest reports, stated that because of AMD's supply constraint due to TSMC issues, they lost desktop discrete marketshare to nV, which is quite interesting, how bad were the yields of AMD's 40nm chips, and imagine how a chip that is ~35% larger would be affected.

It's not the 40nm shortage that caused it, it's the fact that nV flooded markets everywhere with their new 40nm lowend, every shop, supplier and whatnot is flooded with those cards.
Also, isn't JPR counting the chips shipped to AIBs rather than actual cards sold on the end market / oem machines?

AMD shipped around 1.6-1.8 million DX11 chips in Q4 (since 2 million mark was reached in couple weeks of Q1), they were higher midrange and highend chips, even if supply was "unconstrained" for those too, I doubt they would have accounted much more percentage wise from AMD shipped chips, perhaps some 26-28% instead of the ~21-24% but that's about it

 

[Razor1]

Which AMD chips are on 40 nm and which chips are supply constrained? Which AMD chips are being sold right now and process are they made on, to my knowledge all AMD chips are made on 40nm now and have been for a while since sometime last year. And flooding the market with nV 40nm chips, not the case. Why would one AIB be supply constrained and loose market share while the other one can flood the market with theirs to cause a market share lose? Only possibility is AMD isn't purchasing production wafers which 2 q's they have the same problem? Doesn't make sense.

and %'s are off.