GF100 evaluation thread

[no-X]

For me the most interesting thing is the geometry performance of the 16 (14/15) polymorph engines.

It boost performance in low resolutions, where the GTX480 beats HD5870 by 1/3, but it doesn't help to higher resolutions, where geometry isn't limiting factor... GTX480 beats HD5870 in this situation by 10% at best...

http://www.hardwarecanucks.com/foru...s/30297-nvidia-geforce-gtx-480-review-15.html

(not a cherry-picked result, many reviews show it clearly)

Is THIS the best feature of Fermi, which delayed the product by 6 months? Seems to me, that nVidia tried to beat ATi on their own ground and failed... > meh for both

 

[trinibwoy]

From the same article: http://www.hardwarecanucks.com/foru...s/30297-nvidia-geforce-gtx-480-review-25.html

Much in like Battlefield: BC2, the GTX 480’s performance in DIRT 2 is absolutely stunning at 1680 and 1920 resolutions but takes a serious kick to the groin when moving up a notch to even higher resolutions. Once again, NVIDIA is aware that this is happening and there will be improvements soon, but that’s not much consolation for early adopters who are sporting 30” monitors.

Not sure if Nvidia is BSing or there really is a driver issue at higher resolutions but the crash is obvious on both the 470 and 480.

 

[chavvdarrr]

Fermi is DX11 done right.
Face it, R8xx dies on any available DX11 test with tesselation
Its same situation as R300 vs GeforceFX, or 1900 vs 7900 - one was better in older games, the other was ruling in newer ones.
Right now, Fermi is the chip which will have longer lifespan

NV "just" need to fix power issues

 

[compres]

Let's just say I am happy I did not stay home to wait for the reviews.

 

[Mindfury]

Face it, R8xx dies on any available DX11 test with tesselation

Not happened in Dirt2 and S.T.A.L.K.E.R. Call of Pripyat

 

[nyt]

Fermi is DX11 done right.
Face it, R8xx dies on any available DX11 test with tesselation

Scroll down there
http://www.hardware.fr/articles/787-6/dossier-nvidia-geforce-gtx-480.html
and tell me if it is that bad on tesellation. Beaten yes, but dies?

 

[kabab]

Are resolutions above 1920 even relevant? If you look at the steam hardware survey they are not even listed and the majority of people are using displays between 1280x1024 - 1920x1200...

 

[Bouncing Zabaglione Bros.]

Comfort that you won't be missing out on anything? Like Carsten said there are lots of reasons people spend their money and they won't necessarily share yours.

I obviously don't speak for them only for myself - that's why I didn't write that I can't see why anyone would want a 470/480. Other people may be able to justify a purchase because of niche requirements, brand loyalty, etc. The extra cost, heat, noise and power for small speed improvements in certain games/resolutions may be something that isn't important for some people, so those factors can be ignored. I personally don't see the benefits outweighing the negatives because those factors are important for me, but YMMV.

 

[Miksu]

From the same article: http://www.hardwarecanucks.com/foru...s/30297-nvidia-geforce-gtx-480-review-25.html

One should be cautious when reading DIRT 2 benches because it seems quite many sites were running the tests in DX9 mode for the GFX 400-series. Guru3D explained it the best.

edit: Though it seems Hardwarecanucks tested the cards with full version of the game.

 

[air_ii]

Meh for me. Because it's not worth upgrading from 5870, which I've enjoyed for almost 5 months now.

 

[no-X]

Not sure if Nvidia is BSing or there really is a driver issue at higher resolutions but the crash is obvious on both the 470 and 480.

Similar relation can be seen on overall results, too:

1680/AA4x: GTX480 is 18% faster than HD5870 1GB
1920/AA4x: GTX480 is 15% faster than HD5870 1GB
2560/AA4x: GTX480 is 12% faster than HD5870 1GB

http://www.computerbase.de/artikel/...480/21/#abschnitt_performancerating_qualitaet

The higher geometry performance seems to boost framerate only in low resolution... neither the additional 512 megs or higher bandwidth can compensate it...

 

[CarstenS]

Since it's
+18%
+18%
+30% (ok, Framebuffer)
even for 8xAA, this might rule out bandwidth as a limiting factor.

 

[Novum]

The higher geometry performance seems to boost framerate only in low resolution... neither the additional 512 megs or higher bandwidth can compensate it...

This is an expected outcome for me. Lower resolution means smaller triangles.

 

[Jawed]

What are the improvements texturing wise? TUs are now part of SMs, clocked higher, but that's that. Are you including unified texture caches in this?

All I'm saying is that with a significant theoretical texturing rate deficit, GF100 is doing very nicely - and I think post-processing type operations (point-sampling, in a lot of cases - or shading pass in deferred shaders) is working nicely too. Though I think I saw a murmer about the AF hit being worse than GT200, somewhere on some random page in some random review.

Right now, honestly, I'm questioning the value of spending much time "analysing" this thing. If it's missing half its TMUs, has ALUs turned off on the top-most SKU and the GDDR5 has come in way slower than it should have done, I think we're looking at something too broken by circumstance to be quibbling over the final balance of the architecture.

So, I simply see signs of something good.

I think all of ati's recent execution prowess/track record is going to be put to a very stern test as they try to climb the distributed setup/raster/tessellation mountain. Not to mention a brand new process due for the end of this year. We'll know if it was just nv fucking up or the new bits are the really hard deal. Having said that, ati has the luxury of spreading out the changes over cypress->hecaton->NI.

I still can't decide if the vague rumours about a major architectural change for ATI are meaningful/reasonable. My bias is towards thinking that AMD will continue making incremental changes: there won't be a big bang any time soon.

Though there is an argument for saying that ~4 years after R600 means a big bang is due.

Jawed

 

[Neb]

I choosed "480 roxxx, 470 is ok-ok" for the lack of "Quite satisfied with 480, 470 is ok-ok". After seeing most reviews and filtering out the bad reviews, taking into account features and card size it seems quite the deal for me this time. Though I will wait for a custom cooler based GTX480. I might also wait for price drop and/or new revision.

 

[CarstenS]

Though I think I saw a murmer about the AF hit being worse than GT200, somewhere on some random page in some random review.

From my tests, I'd say, AF hit's almost constant compared to GT200 percentage wise. When running the infamous Villagemark in the 2-minute-benchmark mode, i.e. realtime and much more repeatable results, i get the following drops:

GTX 285 (Quality 1:1 AF -> 16:1 AF): 75,75%
GTX 480 (Quality 1:1 AF -> 16:1 AF): 74,60%

GTX 285 (High Quality 1:1 AF -> 16:1 AF): 69,53%
GTX 480 (High Quality 1:1 AF -> 16:1 AF): 69,60%

That's showing two things for me: First, they do not seem to have touched filtering quality on a global scale and second they didn't improve much on the efficiency front in terms of conventional texturing.


BTW:
Where's the rumor with the "missing TMUs" coming from?

 

[Kowan]

Meh for both.
I don't feel a need to upgrade my Tri-SLI 285 setup yet.

 

[Jawed]

BTW:
Where's the rumor with the "missing TMUs" coming from?

The supposed configuration of GF104 (64 TMUs but 256 ALUs), plus the noises that GF100 lost half, which weren't Neliz-only originated - though tracing any rumour back beyond the points of arrival on B3D is tricky

Pretty shaky ground, I admit. Gotta wait for other Fermi GPUs to appear...

Jawed

 

[no-X]

This is an expected outcome for me. Lower resolution means smaller triangles.

...and geometry load is constant (same for 1680 and 2560). That's why I don't like this approach very much. For a 16-cluster chip, which should perform best at 2560, performance is boosted in low resolutions. And for hypotetical mainstream 5-6 cluster chip, which would profit from additional performance in low resolution, geometry performance is cut down

in fact HD5970 with its two triangle setups and two tessellators performs better in many situations, than Fermi with 4 triangle setups and 16 (/15) tessellators

 

[green.pixel]

Meh for both for me.

I think I'll wait for refreshes on both side at the end of the year, if there are some, and in the meantime build some mATX/SSD goodness.

With the way games industry is going on now, there isn't a SINGLE game on the horizon for the next full 1-1.5 year, at least not announced yet, which 1) interest me in a way that 2) it would make me WANT to own one of these cards. All multi-mon/3D/PhysX bullshit aside, just plain good ol' games. Sure, there would be Crysis 2 and Rage, but sure as hell I can't see WHY it couldn't work on current cards just fine and why those games would be in a need of such powerful HW, especially considering the fact that Rage should work on consoles at 720p/60. C2 should IMO be at least ~30-40% faster than the first game at the same res/AA, given the city setting and the abbility to use occlusion culling. Unless, of course, the do it the old way - make a deal with nV/ATI, and intentionally slow down a game so the sheep-users upgrade in an exchange for some cold cash.