NVIDIA GF100 & Friends speculation

[Mintmaster]

If those Chinese numbers are right for HAWX, then I think it's very likely that Fermi sets up and rasterizes multiple triangles per clock, as Razor1 alluded to a little while ago. That game is heavily setup limited.

This also bodes well for tesselation performance.

 

[gamervivek]

Well for one, if their salvage part is faster than Cypress, nothing bodes better for them. Second, since the release is still aligned for March, they may not want to show what their "best" card can do and leave AMD still not knowing what they need to counter, in terms of performance.

Wow, I wish I had that kinda optimism about my looks.Talking in absolutes, you got Charlie beat hands down.

As for felix's numbers, they were done in conditions that we can't compare with these Fermi benchmarks, simply because we don't know the extra hardware involved. So let's wait and see for some benchmarks where the all the conditions are known.

The numbers line up pretty well from the reviews.
Then again what guarantee we have that both the machines in that video had the same hardware?
Or do you want to imply that 5870 might overrun the fermi if paired with more CPU Mhz?
And I believe that was a gtx380.

 

[Mize]

I think it was a Voodoo 5 6000 with optimized drivers

 

[BRiT]

Mize, you're close. It was the 3dfx Rampage with Sage!

 

[Mize]

Mize, you're close. It was the 3dfx Rampage with Sage!

Oooh! That makes me all tingly for some reason

 

[bahanime11]

The FC2 numbers are different because the way the bench was done. Ranch Small is not often used by review sites, who typically fraps a playback scene.

For similar comparison, http://www.pcper.com/article.php?aid=820&type=expert&pid=6

http://www.pcper.com/images/reviews/820/farcry2-1920-bar.jpg

http://www.pcper.com/images/reviews/820/oc/farcry2-1920-bar.jpg

nV gf360 ~= 5870

 

[3dcgi]

If those Chinese numbers are right for HAWX, then I think it's very likely that Fermi sets up and rasterizes multiple triangles per clock, as Razor1 alluded to a little while ago. That game is heavily setup limited.

This also bodes well for tesselation performance.

I haven't paid much attention to HAWX. Why do you think it's heavily setup limited? I would expect a flight game to have a lot of big triangles and "empty" screen space.

 

[gamervivek]

The FC2 numbers are different because the way the bench was done. Ranch Small is not often used by review sites, who typically fraps a playback scene.

For similar comparison, http://www.pcper.com/article.php?aid=820&type=expert&pid=6

http://www.pcper.com/images/reviews/820/farcry2-1920-bar.jpg

http://www.pcper.com/images/reviews/820/oc/farcry2-1920-bar.jpg

nV gf360 ~= 5870

Thanks, those settings are the closest to the leak, still can't be sure about 16xAF and if the other card was gtx285.

 

[DemoCoder]

I haven't paid much attention to HAWX. Why do you think it's heavily setup limited? I would expect a flight game to have a lot of big triangles and "empty" screen space.

I thought flight sims have always been somewhat geometry limited, because there's very little of the traditional occlusion culling data structures used (BSPs/portals/etc). Typically, you have an extremely large view distance and you're only going to be using large flat triangles if you're not modeling lots of mountain terrain and ground shrubbery.

I remember in the old days of B3D, the sims were always CPU liimited (prior to the era of T&L/vertex shaders) because of the geometry load.

 

[Mize]

The FC2 numbers are different because the way the bench was done. Ranch Small is not often used by review sites, who typically fraps a playback scene.

For similar comparison, http://www.pcper.com/article.php?aid=820&type=expert&pid=6

http://www.pcper.com/images/reviews/820/farcry2-1920-bar.jpg

http://www.pcper.com/images/reviews/820/oc/farcry2-1920-bar.jpg

nV gf360 ~= 5870

?

I have a 5870 and it benches on the FC2 bench tool at 68 with an OC'd Q9550 - nowhere near the fraps ones you've linked to.

 

[gamervivek]

?

I have a 5870 and it benches on the FC2 bench tool at 68 with an OC'd Q9550 - nowhere near the fraps ones you've linked to.

They are the ranch small benchmarks, not fraps one, as posted in the leaked video.
And far cry 2 likes i7 proccies, put together with the cleaner install/different OS of the review site and hence the discrepancy in your results and theirs?

 

[Mize]

They are the ranch small benchmarks, not fraps one, as posted in the leaked video.
And far cry 2 likes i7 proccies, put together with the cleaner install/different OS of the review site and hence the discrepancy in your results and theirs?

That didn't look like fraps unless I'm confused. It had the three cycles and average like the FarCry2 bench tool.

 

[DemoCoder]

If those Chinese numbers are right for HAWX, then I think it's very likely that Fermi sets up and rasterizes multiple triangles per clock, as Razor1 alluded to a little while ago. That game is heavily setup limited.

What do you think about running the setup in its own clock domain (e.g. hot clock) rather then using multiple units? That would seem like the simplest way to do it.

 

[jimmyjames123]

http://www.youtube.com/watch?v=XFgzG9WyGNQ

Although this raytracing demo shows GF100 with 3x performance improvement over GTX 285, I wonder when we will be able to see anything remotely close to this in a real game. Probably many years off

 

[ninelven]

Only about 10 years.....

 

[Silus]

Wow, I wish I had that kinda optimism about my looks.Talking in absolutes, you got Charlie beat hands down.

Right

The numbers line up pretty well from the reviews.
Then again what guarantee we have that both the machines in that video had the same hardware?
Or do you want to imply that 5870 might overrun the fermi if paired with more CPU Mhz?

Not exactly. even the pcper numbers are way off than for example, Tech-Report, which uses a more or less similar system, with a Core i7 965 Extremne (albeit without the "slight" overclock to 3.33 Ghz)

http://techreport.com/articles.x/17986/6

67 fps average @ Very High (not even Ultra high). A "far cry" from ~84 fps of pcper. Some differences in the system and even the exact place where the benchmark took place, will cause discrepancies, but whatever they are + the 133 Mhz overclock on the CPU, wouldn't cause a boost in fps of over 20% for the HD 5870. And even if it did, the effect could be the same for the Fermi based system. But we don't know that yet. So let's wait for benchmarks where all the conditions are known.

If I go by tech reports numbers, that also include a GTX 285, the non-Fermi based system seems to be using a GTX 285.

And I believe that was a gtx380.

We'll see.

 

[MarkoIt]

Although this raytracing demo shows GF100 with 3x performance improvement over GTX 285, I wonder when we will be able to see anything remotely close to this in a real game. Probably many years off

0.6 fps... we need a 50 fold improvement for 30fps.
If they can deliver a 2x improvement every years, in 6 years we will have it running in realtime.
If they deliver a 2x improvement every 18 months , it will be 9 years.
By 2020 we will probably able to render Avatar in realtime
I guess we are going to see TWIMTBP's logo before the beginning of every CG movie

 

[PSU-failure]

The raytraced scene itself seems too light.

0.6fps is quite good for heavy raytracing, but without any indication on the scene itself (number of lights?) and the recursivity level (which seems rather low) it's just plain useless.

What about the same scene on Cypress? It's not that hard to port the RT algorithm to DC or OCL and in fact efficiency would be high given the fact RT uses many Vec3/Vec4 data, but wait... it's a proprietary algorithm!

It's always the same, they show proprietary stuff again and again, so that we can't compare to anything else (except perhaps very bad CPU implementations such as with PhysX).

 

[OlegSH]

The FC2 numbers are different because the way the bench was done. Ranch Small is not often used by review sites, who typically fraps a playback scene.

For similar comparison, http://www.pcper.com/article.php?aid=820&type=expert&pid=6

http://www.pcper.com/images/reviews/820/farcry2-1920-bar.jpg

http://www.pcper.com/images/reviews/820/oc/farcry2-1920-bar.jpg

nV gf360 ~= 5870

At first you need to look at what exactly test they push at video, this is RANCH SMALL(also has RANCH MIDDLE and LARGE), then try to look at FPS of second card it's 50FPS, then look at pcper reviews with R5870 and GTX285, GTX285 at same 1920*1200 16AF 4AA has 65 FPS, so this not RANCH SMALL and you just compare apples with oranges
Then look back at http://www.pcgameshardware.com/aid,...test-DirectX-11-graphics-card/Reviews/?page=4 , try to look GTX285 FPS and processor spec, so if we compare FPS of gtx285 from video and pcgameshardware difference only 3FPS so we can count'em as processors difference FPS, all others reviews what i seen show FPS of stock clocked R5870 at ranch small at same options as 64FPS, so if try to compare with GF100 FPS it's 84/0.64=131 or 31% faster, And about tested GF100 card, it's not GTX380 or 360 cause much of things not yet finalizing

 

[DemoCoder]

What about the same scene on Cypress? It's not that hard to port the RT algorithm to DC or OCL and in fact efficiency would be high given the fact RT uses many Vec3/Vec4 data, but wait... it's a proprietary algorithm!

Too much hand waving. Not that hard to port to DC/OCL? A trivial port might not be that hard, but an optimized port? If it were so easy, many people would have already done it. Hell, if I thought it were that easy, I'd do a weekend port and start selling it to Autodesk, et al. You can't just hand wave away efficiency issues because there's a lot of vector processing. Much of the cost of ray-tracing is in building and traversing spatial acceleration data-structures, and this is an area where memory/cache architecture matters more than raw ALU power. It could very well be that for this workload, Fermi could be much faster.

Also, I don't see what's wrong with TechDemos. They're trying to show what's possible, and NVidia isn't the only company to show off tech-demos at trade shows.