NVIDIA Fermi: Architecture discussion

Actually looks a lot more stable with one large spike in the middle.

 

With moving camera trough the scene your frame rendering speed is changing every frame depending on the frame complexity so the graph is very far from spikey.
In fact if u watch closely, both hd5870 and gf100 are copying each others frame risses and falls nicely so theres no way hd5870 is any way more limited by triangle setup than GF100 (at least in uniengine heaven).

 

True but wasn't this segment from a heavy tessellation part of the demo? You won't see much variation if this slice of the benchmark was consistently geometry heavy. We need to compare geometry light and geometry heavy sections to see whether or not GF100 handles the latter better than Cypress.

 

Look at the seconds 23 - 28. 5870 jumps from 26 fps to 31 fps, GF100 from 36 to 57. The 5870 is limited by something else than the GF100.

 

Nowhere in history has a canned benchmark from a vendor been so profoundly and throughly studied like this one. Is like the holy grail from the dark holes from nvidia and yet, you guys study it like if there is no tomorow.

The fact that fermi is not coming in a couple of months aggravates the situation.

Damn you guys are desperate

 

Increasing framerates such disproporcionaly to the rest is not a big win (but it can help u score better average fps in charts ) It can also be something with scalar vs vector sp-s or the new shared L2 cache on gf100. Both cards rise fps there in the graph and thats the point.

 

i would say 1.6 is about equal performace

334 X 1.6 = 534.4



i know its rather meaningless but still food for thought.

 

Thankfully unlike with DX10 we won't need to rely on synthetic tests this time around. By the time Fermi ships Dirt2, AvP, Stalker and (I think) BF:BC2 will all serve as DX11 benchmarks.

 

Conversely during secs 32 - 37, Cypress shows a jump of 50% where as GF100 does only ~12%. Taking apart a vendor provided bench isnt some thing scientific when the whole testing process itself cannot be verified.

If Fermi is on average 1.6 times faster then Nvidia would have finally achieved the perf/mm2 of AMD. That would be a big win.

 

Good point... especially since that means 1/2 and 1/4 size derivatives would be very compelling for medium and low end parts.

 

Let's keep this in perspective here, we're just looking at one portion of a tesselation benchmark.

If it was 60% fairly consistently in games and such then yes, it would be a big win.

 

I doubt thats the case though. They are within ~10% when they start the 60 second clip, and at the end of the clip the ATI card is climbing back up hard, I suspect in the full bench they are within 15-20% of each other.

 

There's a difference for me between it looks like X% today and what the benchmarks from a large variety of games (and varying settings) from independent sources will show after the GF100 launch.

The real research and personal evaluation can only start after the latter for me. There I can have a better look at price/performance ratios and get a few very valuable hints on image quality.

And frankly the last is for the moment my biggest concern; I've always been extremely picky with AF quality and extremely allergic to a crapload of AF related optimisations. I can only hope that the low amount of TMUs won't mean that between default driver settings and "high quality" mode the performance difference will be as small as since G80. In a worst case scenario where the performance drop might be a flashback to prior than G80 solutions I'm afraid it will degrade significantly the overall value of a GF100 for me.

 

Agreed. Fortunately I suspect this is highly unlikely, though at the same time it's a bit worrying that they didn't even touch on the anisotropic filtering quality in the press releases.

 

Keep an eye on section 4.0:

http://developer.amd.com/gpu/ATIStreamSDK/assets/ATI_Stream_SDK_Getting_Started_Guide_v2.0.pdf

when the Evergreen FireStream cards arrive. I don't expect Juniper based FireStream to support double precision.

Jawed

 

That would be odd, whats the point of a firestream card if it can't do DP? The whole thing doesn't make much sense to me, the power/perf and perf/cost of 57XX compared to 58XX doesn't really make it a much, if any better option. In fact, in power/perf is lower to use 57XX.

 

Let us hope you are right.

 

Not in the press releases but TR got a related quote. TR is referring to a cleaner AF test pattern on Cypress but in motion it's been demonstrated that Nvidia's AF is still superior (less shimmering) so unless they've gone backwards it should be ok.

 

That's more or less what I expected. Works for me.

 

Considering usable DP support didn't exist at all on professional cards a few years ago (and it could be argued that RV770 and GT200 implementations weren't useful either) then I don't see why that should be a problem, especially for lower end boards. If companies want the best they need to pony up the $$$.