First few GFX benches

Status
Not open for further replies.
O

Ollo

Newcomer
Found this while browsing rage3d:

http://www.maximumpc.com/features/feature_2003-01-03.html

"Dare to Compare: GeForce FX Early Benchmarks

GeForce FX

Quake3 Demo001, 1600x1200 2xAA: 209fps

UT 2003 Asbestos, 1600x1200 2xAA: 140fps

3DMark Game4, 1600x1200 2xAA: 41fps



Radeon 9700 Pro

Quake3 Demo001, 1600x1200 2xAA: 147fps

UT 2003 Asbestos, 1600x1200 2xAA: 119fps

3DMark Game4, 1600x1200 2xAA: 45fps

Tests were run in the Alienware prototype system."
 
In what might be a first, The actual "write-up" surrounding those benchmark numbers makes almost complete sense. Maximum PC actually demostrated that they have some sort of clue about R300 and NV30 architectures, and bottlenecks. That impresses me more than the scores themselves.

GeForce outperforming Radeon at high resolution and low AA (and no aniso) settings is predictable. This is a relevant quote:

Our hunch is that turning on 4x anti-aliasing at 1600x1200 would diminish the GeForce’s performance lead over the Radeon, or maybe even nix it entirely. But that’s just a guess based on the scores we achieved, and the fact that nVidia wouldn’t let us run anything that would stress the memory pipeline.
Click to expand...

One would start to get really suspicious if nVidia only allowed 16 bit scores. ;)

Like MaxPC says though, the relatively low 3DMark Game 4 numbers are indeed a surprise.

Though MaxPC isn't completely in the clear...they seem to be on the wrong track about ATI's answer to NV30:

This is just the first round, though. We have no doubt that ATI has plans for a souped-up Radeon that will be ready to roll as soon as the GeForce FX ships. And if you really twisted our arms, we’d bet money that it will be running on a 0.13-micron core and using 256-bit DDR II memory.
Click to expand...

In reality, it's just short of being an official fact that R350 is 0.15 micron, and will likely not use DDR-II.
 
They say that game 4 uses heavy vertex and pixel shaders. I would have thought though that on r300 and nv30 the shaders involved don't tax either card and the big strain in this test is the leaves. It'pretty much a single texture fillrate test and so could easily be heavily bandwidth limited on an unbalanced card like the FX.

I would think that that test just comes down to r300s 19.8gb/s v nv30s 16gb/s and this is why the r300 is ~10% ahead.
 
I would think that that test just comes down to r300s 19.8gb/s v nv30s 16gb/s and this is why the r300 is ~10% ahead.
Click to expand...
Exactly! :)
The other results were pretty much expected right from the start, and with further optimizations, they would only increase...
 
IMO, this is more of a board-related thing than a technology-related thing... So I guess we better talk about it there.
Click to expand...
Frankly, that doesn't make much of a difference... e.g, we could talk about that in both forums :) only if there was something to hold a discussion about, as the results are very much predictable and as far as I know, the sample isn't final anyway...
 
Bambers said:
They say that game 4 uses heavy vertex and pixel shaders. I would have thought though that on r300 and nv30 the shaders involved don't tax either card and the big strain in this test is the leaves. It'pretty much a single texture fillrate test and so could easily be heavily bandwidth limited on an unbalanced card like the FX.

I would think that that test just comes down to r300s 19.8gb/s v nv30s 16gb/s and this is why the r300 is ~10% ahead.
Click to expand...

Well for starters I question whether the benchmarks are for real, but for the sake of the argument I'm not so sure it comes down to difference in fillrate and memory bandwidth.

I always thought that Nature was mainly about vertex shader performance (with all those moving leaves) and so I would suggest that the special array of units in NV30's vertex shaders wouldn't be optimized yet. Besides, it shouldn't be to difficult to hold most of the textures for those leaves in a cache - so that's another reason to dismis the memory bandwidth idea. :|
 
I'm also partially doubting the benchmarks. Seems a little odd to me, normally you suddenly get reviews popping up on tomshardware, anandtech and hardocp within minutes of each other, not some 3 benches in a magazine :-?

Still I did a test with my 8500 on xp2100+

Core was set to 300Mhz and I ran nature (repeat set to 2) and got

51.2fps with ram at 325Mhz and 42.5fps at 250Mhz
 
Bambers said:
I'm also partially doubting the benchmarks. Seems a little odd to me, normally you suddenly get reviews popping up on tomshardware, anandtech and hardocp within minutes of each other, not some 3 benches in a magazine :-?
Click to expand...
With early silicon? With NVIDIA being so restrictive on the benchmarks used and on testing in general? I don't see any reason whatsoever why these benchmakrs would be suspicious or anything of this sort...
 
Okay, so I'll continue discussing this here... But it would be good if a mod locked one of the threads, dividing a discussion is never good.

LeStoffer: Err, what? Holding "most of the textures of those leaves in a cache" ? What are you talking about? There's no huge caches on a GPU. Yes, there is a texture cache. But it certainly couldn't handle all of a texture! It only handles part of a row of a triangle. Nothing more.

As for Nature being Vertex limited. I'm not sure, but no matter what, more vertices ALSO means more memory bandwidth is required ( more complex vertices, however, don't )
That's because transformed vertices are stored in memory before using the indices to send some of them to Triangle Setup.


Uttar
 
Uttar said:
LeStoffer: Err, what? Holding "most of the textures of those leaves in a cache" ? What are you talking about? There's no huge caches on a GPU. Yes, there is a texture cache. But it certainly couldn't handle all of a texture! It only handles part of a row of a triangle. Nothing more.
Click to expand...

Sorry for being a bit sloppy with the way I put it. :arrow: I meant the say that the textures for those the leaves are so relatively small that the card shouldn't need to resort to any heavy texture fetches from memory to render the pixels in that operation. The texel data should be able to fit efficiently in a cache thus masking any memory bandwidth difference out.

But then again: Only Futuremark knows how efficient their game engine/ render is. If they wanted it to depend on memory bandwidth over fillrate they could of course have done just that.
 
So I assume actually meaningful benchmarks are due in short order, maybe graced with a screenshot comparison somewhere...?

Can any reviewer type person drop any extremely vague and annoying hints along those lines...?

BTW, are these cool :!: emoticons new, or is my memory just fuzzy today? :?:
 
alexsok said:
Exactly! :)
The other results were pretty much expected right from the start, and with further optimizations, they would only increase...
Click to expand...

Well, let's see how things stack up at 1280x1024 and 1600x1200 with 4x AA and 8-16x AF before getting too giddy and assuming only one IHV can improve performance through "further optimizations".
 
Hrm, I am of the mind that these are benchmark numbers are fairly scetchy. I will echo what someone already else said, there ought to be some in depth reviews of the hardware at Toms, Anand and the like before we can draw conclusions. Further it appears these are the same performance numbers that are in this article were released from nvidia some time ago and the writers simply used existing benches of the Radeon 9700 pro to compare at same resolution and AA levels.(please correct me if I am wrong here.)

*clears throught* There have been no in depth reviews of the hardware even though the card was launched in mid November. Nvidia is clearly still having difficulties with the product production otherwise we would be seeing pre shipping samples at reputable review sites in order to stifle sales of the Radeon 9xxx series cards. The nv30 is rumored to be on sale from certain third parties within the next two weeks. A rumor that I would question heavily in lite of the lack of pre hype reviews it would certainly be a lack luster launch. I won't expect to see the nv30 shipping in any sort of volume for at least a month after we start seeing more in depth reviews, much the same way the Radeon 9700 pro was. I am guessing we will see reviews hopefully by the end of the month with volume shipping sometime mid to late February.
 
Uttar said:
As for Nature being Vertex limited. I'm not sure, but no matter what, more vertices ALSO means more memory bandwidth is required ( more complex vertices, however, don't ). That's because transformed vertices are stored in memory before using the indices to send some of them to Triangle Setup.
Click to expand...
Hold on a minute here!
Why the hell would GeForceFX need to store transformed vertices back into video memory and then read them again at triangle setup stage?? The ONLY think this could be good for is their render to vertex array functionality, but you don't need to do this GENERALY in every case!? Where did you get this info from (and how sure are you about it)?
 
The nature test is more about bandwidth than Pixel / Vertex Shader speed because of all the Alpha textures being used (search the forums). Its not much surprise R300 is out performing there considering this.

Given that the AA pipeline was 'optimised for 4X AA', if these are real benchmarks I find it surprising that 4X wasn't used. If this is the case then 4X will be quite a telling mode. 6X to 6xS comparisons will be telling as well.
 
Well, let's see how things stack up at 1280x1024 and 1600x1200 with 4x AA and 8-16x AF before getting too giddy and assuming only one IHV can improve performance through "further optimizations".
Click to expand...
You're right there, but I have a hunch that even there, GeForce FX might come out a winner... but the testing settings you described particularly stress the memory bandwidth, something NVIDIA doesn't want to show to anyone yet, since that'd show the GeForce FX in a dim light...

Still though, it would be fairly interesting to see how the events unfold in the near future.
 
DaveBaumann said:
Given that the AA pipeline was 'optimised for 4X AA', if these are real benchmarks I find it surprising that 4X wasn't used. If this is the case then 4X will be quite a telling mode. 6X to 6xS comparisons will be telling as well.
Click to expand...

Indeed. What is curious, if bandwidth is the culprit on why they used 2x instead of 4x, wouldn't we have expected them to dial the resolution back to 1280x960 and still use 4x? Given that Nvidia picked all the details on the test, I mean.
 
DaveBaumann said:
The nature test is more about bandwidth than Pixel / Vertex Shader speed because of all the Alpha textures being used (search the forums). Its not much surprise R300 is out performing there considering this.
Click to expand...

Okay (sorry for going bit off topic)...

In this thread: http://www.beyond3d.com/forum/viewt...order=asc&highlight=alpha textures nature - which Dave is probably refering to - Nicklas "worm" Renqvist give this info:

Rendered triangles per frame (min/avg/max): 55601/81722.5/180938

Rendered textures per frame with 16 bit textures (min/avg/max): 14.9/17.4/20.7 MB
Rendered textures per frame with 32 bit textures (min/avg/max): 28.4/33.4/40.0 MB
Rendered textures per frame with texture compression (min/avg/max): 10.4/12.0/14.6 MB
Rendered textures per frame with DXT1 compressed textures and 32 bit transparent textures (min/avg/max): 13.0/15.0/17.8 MB

While this is per frame does that really translate into a memory bandwidth limited situation?
 
Status
Not open for further replies.
Back
Top