AnandTech - GeForceFX 5800 Ultra..... Canceled
- Thread starter Sabastian
- Start date
OpenGL guy
Veteran
Ahem... CPU limited. The 9700 Pro system was reported as a 1.3 Ghz P4 vs. 2.5 Ghz P4.DaveBaumann said:
Mintmaster
Veteran
Those are the scores from the old drivers, aren't they? A GFFX 5800 Ultra gets over 5000 points, so a 400 MHz version should be able to get over 4000 points.
Sabastian said:
The greatest speed gains are in GT2 and GT3, which don't use PS 2.0 (i.e. floating point) at all. There are large gains in GT4 as well, to be sure, but considering how small a fraction of it uses FP shaders I doubt FP16 vs. FP32 has very much (if anything) to do with the performance gains.
Althornin said:
In general I would agree, but as GFfx isn't available to consumers yet I'm not really bothered by some slight rendering problems which as you say probably don't significantly change performance anyways. The point of GFfx benchmarks right now is to get a sense of what the architecture is capable of and to predict how it will match up to e.g. 9700 Pro. Using those criteria, it seems silly to use the old scores (which obviously have almost nothing to do with what actual performance will be) just because the new scores are not yet on compliant drivers. OTOH, the new scores should only be used on a provisional basis; they are obviously not "official" scores.
Still, I think there's a difference between a situation where, say, leaked beta drivers for an established GPU give a 5% performance boost but introduce rendering errors which may be responsible for the entire speedup, and the current situation, where all drivers are beta, and the new version gives a 100% speedup, only a small portion of which might be attributable to the rendering errors.
Thanks for the input Dave. Also there is the matter of slower memory on top of the lower clocked core. Just what is a more accurate score for the non ultra anyhow? I can't find any results with the drivers in question. I think Dave B has the non ultra version of the card, maybe he will shed some light on the matter. Again, thanks for the heads up with regards to the drivers in question.
Mintmaster
Veteran
Dave H, I agree with your point in that we are interested in what the architecture is capable of. That's why I mentioned the new drivers.
However, Sabastian does have a point, although its not related to FP16 vs. FP32. These gains are suspect because they look like a case of hand tuned optimization and instruction scheduling by the NVidia driver guys, as we haven't seen any other shaders improving in performance. A case in point is the digit-life review, where all their shaders run slow on NV30 even with the new drivers.
Obviously NVidia would go to some lengths to get these shaders optimized, and even so NV30 does not perform as well as R300 in shading, even with 50%+ higher clock rate and more transistors.
The point is I don't see how compiler technology is going to improve the situation enough, and even if it does, I think what we are seeing with 3DMark2k3 is an upper limit. If you write an application with any shaders more complex than PS 1.1, chances are it will run like crap on NV30 unless you are a major developer with games that get benchmarked on major review sites.
However, Sabastian does have a point, although its not related to FP16 vs. FP32. These gains are suspect because they look like a case of hand tuned optimization and instruction scheduling by the NVidia driver guys, as we haven't seen any other shaders improving in performance. A case in point is the digit-life review, where all their shaders run slow on NV30 even with the new drivers.
Obviously NVidia would go to some lengths to get these shaders optimized, and even so NV30 does not perform as well as R300 in shading, even with 50%+ higher clock rate and more transistors.
The point is I don't see how compiler technology is going to improve the situation enough, and even if it does, I think what we are seeing with 3DMark2k3 is an upper limit. If you write an application with any shaders more complex than PS 1.1, chances are it will run like crap on NV30 unless you are a major developer with games that get benchmarked on major review sites.
TheMightyPuck
Newcomer
Still, for better or worse, we should wait until the GF:FX makes it to market (if ever) before we categorize it absolutely. Damn, look at me, I almost feel sorry for the little chip.
This is a very good point. Although, just to play devil's advocate for a moment, it's worth pointing out that while the Digit Life review--with new drivers--showed GFfx losing to 9700 by a factor of ~2 on the Rightmark shader tests, back a few weeks ago Brent posted on this forum some Shadermark results--with the old drivers--that showed GFfx losing by a factor of ~3. (edited for Tagrineth) Is it the new drivers, or the difference in tests? I dunno.
Still, there are explicit comments from Nvidia engineers to the effect that the new drivers hand-optimize for the 3DMark03 shaders, if not by actually optimally scheduling the shaders in the drivers themselves, at the least by deciding case-by-case whether to run a PS 1.4 shader in NV30's general-purpose PS 1.4-2.0+ shader pipeline or to break it up into several passes through the legacy PS 1.0-1.3 path.
Now, the first thing to be said is that these statements are obviously part of Nvidia's campaign to discredit 3DMark03. They're probably true, but just as clearly Nvidia (and ATI) do the same sort of thing for previous 3DMark benchmarks, and for any games heavily used by benchmarking sites; the only difference is that they don't proclaim it so publicly and in a way that suggests the tests themselves are invalid.
The second thing to note is, golly, it sure doesn't say good things for NV30's real shader pipeline if emulating a PS 1.4 shader with 2 or 3 PS 1.1 passes is faster in many cases.
Now, with those two things out of the way I think it's obvious that normal games (i.e. outside of the 4 or 5 somehow annointed as benchmarks) are not going to get special case treatment in the drivers. On the other hand, it's also obvious that developers are going to optimize their games for Nvidia GPUs. If it's "just" a matter of trying the same effect in both PS 1.4 and multipass PS 1.1 configurations, and putting in a special path for Nvidia cards to do whichever is faster, most developers are going to do it (especially as they will have help from Nvidia devloper relations). So in some sense, it's not really that unrepresentative if the special optimizations made in the drivers are likely to be done by developers for most games.
If it's a matter of explicitly scheduling the shader in a way different from what the general-purpose compiler in the driver would do, then of course developers aren't going to have the ability to do that on their own. I find it difficult to believe that NV30's shader pipeline is so difficult a compiler target that future drivers won't be able to at least approach the performance of hand-optimized code. Still, I'd much rather wait for drivers that optimize in the general case than see hand-tuned results right now.
In case I sound inconsistent: the difference between the two cases is that in the first the developer has the ability to do the optimization themselves and likely will, but in the second they can't (unless perhaps if they use the proprietary NV OpenGL extensions).
Mintmaster
Veteran
Dave H said:
I think its kind of funny that you see losing by a factor of 2 rather than 3 as sort of a positive thing
On the other hand, Ilfirin wrote some long shader tests recently and GFFX was losing by a factor of over 7. I would definately attribute your observation to variation in tests rather than improvement in drivers.
Personally, I think dependant textures are the major culprit in NV30 crappy shader performance. The latency problem mentioned by DaveB also fits in with that theory, as you have to buffer out the latency of the texture fetch before you can proceed with executing the shader. This is probably the issue with Xabre's pixel shaders as well.
This is in one of the ATI optimization docs:
It looks like ATI did so well with R300 because it learned many lessons about general dependent texturing from R200. NVidia's dependent texturing in GF3/GF4 was quite fixed and predictable, and JC described it as just an extension of the register combiners.
NVidia also mentioned somewhere that it has no limit on levels of dependancy in NV30, although I don't see why such a feature is useful, as very few shaders go beyond even one level. It may be, then, that they did not optimize dependent textures at all to allow such a capability, and NV30 pretty much just stalls until the texture fetch is complete.
Mintmaster said:Dave H said:
I think its kind of funny that you see losing by a factor of 2 rather than 3 as sort of a positive thing
He said the FX outperformed the 9700 2:1 in RightMark3D, but that with Shadermark the FX was outperformed by the 9700 3:1.
PARALLEL SENTENCE STRUCTURE, PEOPLE! PLEASE!
Similar threads
- Deleted member 13524
- Site Feedback