NV35 Reviews
- Thread starter BenSkywalker
- Start date
Anandtech:
OK, he starts off by mentioning how nVidia called 4 pixel pipelines 8, but then goes on to dictate that ATI should be blamed just as much for not being forthcoming. But he doesn't mention for what particular misinformation ATI is being brought up.
Then he has a creditable conceptualization of the term pipeline, except that he proposes it as a pretext to minimize the significance of the idea that processing can be limited in the end by the actual hardware (and to further promote the idea that ATI and nVidia are equivalent in the secrets they keep because neither provides specifical low level details)... and then to further go on to interject nVidia's completely unsupported comments about their own superiority in this "mystical" practice that seems to supercede the actual hardware capabilities.
Astounding lack of integrity, intermixed with a creditable explanation and established willingness to criticize nVidia to sell it better, but the flaws in it are independently evident and inappropriate, and all his prior behavior establishes is that an (immutable) vendor bias doesn't seem to be the reason for them.
A pretty blatant spin to start off with, and then a technical example leading up to a pretty dedicated obfuscation of the term "pixel" by such terms as "Z pixel", "texture/clk", "Stencil pixel", and "pixel shader ops", where "Color + Z pixel" is established as the special case that doesn't weigh as heavily (it's only listed once, after all, whereas the rest of the items present a count of "8" for them that are more numerous). Nevermind that they are the special cases of splitting Color+Z up into its parts, and that the Color+Z limitation is the most frequent one, even with pixel shading.
Notable quote: "What the marketing folks have done to help clear up the confusion is come up with a list of scenarios and the throughput of their GPUs in those scenarios". <-You're in trouble when the nVidia marketing people are "clearing up" technical discussion for you.
Another: "ATI's R3xx architecture differs from NVIDIA slightly in this respect, as they are able to output 8 pixels per clock in every one of the situations listed above. The advantage isn't huge as it is mostly limited to older games, but the difference does exist and is worth pointing out.".
Now, this is the type of review I think Rev had in mind.
Further confusion: "On ATI's R3xx GPUs this means that there are clusters of 4 x 24-bit FPUs, while on NVIDIA's NV3x GPUs there are clusters of 4 x 32-bit FPUs. All operations on these FPUs occur at the same latency, regardless of precision. So whether you're doing a 16-bit add or a 24/32-bit add, it occurs at the same rate."
Ah, so the R300 is only 4 too, that's why they are just as bad as nVidia? Well, that's not technically what they said, since they didn't mention exactly how many clusters the R300 had, and didn't actually rule out scalar/float3/tex op combinations. Of course, it is implied that they are talking about NV35/30 versus R350/300 (does the NV31/34 have 4 x 32-bit FPUs for the fragment processing under discussion?), so it would have been helpful if their implications weren't so slanted.
The aniso comparison is incomplete, and seems to be a a bit shallow as a sop to prevent ATI from thinking too negatively of Anand following the BS zone the reader had to pass through to reach it.
It would be nice if all the parts of a review were equally unbiased instead of these games of "back and forth biasing" in one article (though it is better than back and forth in successive articles, I suppose), but I guess he wanted to mend some fences, while still "throwing bones" to readers looking for "hard sell" alarm bells and to ATI (and ATI fans) looking for "bias" alarm bells.
His practice of mentioning the competition's expected response does seem to remain a universally applied closing tactic, however.
OK, he starts off by mentioning how nVidia called 4 pixel pipelines 8, but then goes on to dictate that ATI should be blamed just as much for not being forthcoming. But he doesn't mention for what particular misinformation ATI is being brought up.
Then he has a creditable conceptualization of the term pipeline, except that he proposes it as a pretext to minimize the significance of the idea that processing can be limited in the end by the actual hardware (and to further promote the idea that ATI and nVidia are equivalent in the secrets they keep because neither provides specifical low level details)... and then to further go on to interject nVidia's completely unsupported comments about their own superiority in this "mystical" practice that seems to supercede the actual hardware capabilities.
Astounding lack of integrity, intermixed with a creditable explanation and established willingness to criticize nVidia to sell it better, but the flaws in it are independently evident and inappropriate, and all his prior behavior establishes is that an (immutable) vendor bias doesn't seem to be the reason for them.
A pretty blatant spin to start off with, and then a technical example leading up to a pretty dedicated obfuscation of the term "pixel" by such terms as "Z pixel", "texture/clk", "Stencil pixel", and "pixel shader ops", where "Color + Z pixel" is established as the special case that doesn't weigh as heavily (it's only listed once, after all, whereas the rest of the items present a count of "8" for them that are more numerous
). Nevermind that they are the special cases of splitting Color+Z up into its parts, and that the Color+Z limitation is the most frequent one, even with pixel shading.Notable quote: "What the marketing folks have done to help clear up the confusion is come up with a list of scenarios and the throughput of their GPUs in those scenarios". <-You're in trouble when the nVidia marketing people are "clearing up" technical discussion for you.
Another: "ATI's R3xx architecture differs from NVIDIA slightly in this respect, as they are able to output 8 pixels per clock in every one of the situations listed above. The advantage isn't huge as it is mostly limited to older games, but the difference does exist and is worth pointing out.".
Now, this is the type of review I think Rev had in mind.
Further confusion: "On ATI's R3xx GPUs this means that there are clusters of 4 x 24-bit FPUs, while on NVIDIA's NV3x GPUs there are clusters of 4 x 32-bit FPUs. All operations on these FPUs occur at the same latency, regardless of precision. So whether you're doing a 16-bit add or a 24/32-bit add, it occurs at the same rate."
Ah, so the R300 is only 4 too, that's why they are just as bad as nVidia? Well, that's not technically what they said, since they didn't mention exactly how many clusters the R300 had, and didn't actually rule out scalar/float3/tex op combinations. Of course, it is implied that they are talking about NV35/30 versus R350/300 (does the NV31/34 have 4 x 32-bit FPUs for the fragment processing under discussion?), so it would have been helpful if their implications weren't so slanted.
The aniso comparison is incomplete, and seems to be a a bit shallow as a sop to prevent ATI from thinking too negatively of Anand following the BS zone the reader had to pass through to reach it.
It would be nice if all the parts of a review were equally unbiased instead of these games of "back and forth biasing" in one article (though it is better than back and forth in successive articles, I suppose), but I guess he wanted to mend some fences, while still "throwing bones" to readers looking for "hard sell" alarm bells and to ATI (and ATI fans) looking for "bias" alarm bells.
His practice of mentioning the competition's expected response does seem to remain a universally applied closing tactic, however.
Joe DeFuria
Legend
saf1 said:
What conspiracy theories? I offered explanations of how NV35 pixel shading benchmarks can show significant improvement (clock for clock) against NV30, without there being drastic core changes. (Which likely would not be possible for a refresh.)
The first explanation was due to raw bandwidth increase.
The second explanation was due to hardware tweaks, just not drastic ones.
The third explanation was software tweaks.
What's your explanation? They threw out the NV30 core and started from scratch? FYI, my guess is a combination of the second and third options.
Within the limits of what they discussed, I think the article was good (unlike the Anandtech issues, which IMO are bad, period). The problem is that those limits currently facilitate expressing misinformation, even when sticking to just the facts, an occurrence I've mentioned before. Since that is somewhat unavoidable without omniscience, and the discipline of benchmarking is consistently applied, I'm criticizing the article, not Mr. Salvator (sp?).Reverend said:
Doom3 can run in fixed-point on the NV3x without any perceptible loss of image qualit, which isn't surprising, since Doom3 was developed with the NV30 in mind as the high-end platform. With all the fixed-point functional units the NV35 has, it does very well on engines written for that architecture. AFAIK the R3xx does not have as many floating-point functional units as the NV35 has fixed-point functional units, so I don't think it is surprising that it can't keep up with the NV35 running code that was designed for it.
If the NV3x success at executing Doom 3 leads to success for the architecture in the marketplace, I think we can wave bye-bye to a vendor-independent shader model for this generation, simply because the DirectX and OpenGL vendor-independent shading models assume floating-point functional units. Developers that want to develop competitive titles that run well on NVidia cards will be forced to use NVidia-specific extensions.
ATI might have to come out with a Cg backend for R300 cards after all.
If the NV3x success at executing Doom 3 leads to success for the architecture in the marketplace, I think we can wave bye-bye to a vendor-independent shader model for this generation, simply because the DirectX and OpenGL vendor-independent shading models assume floating-point functional units. Developers that want to develop competitive titles that run well on NVidia cards will be forced to use NVidia-specific extensions.
ATI might have to come out with a Cg backend for R300 cards after all.
BenSkywalker
Regular
Correct. However, and obviously I can't speak for Dave Salvator since I can't know his schedules, there are political and diplomatic behind-the-door issues to resolve prior to making any potentially, uh, "smoking gun" issues public, and given the nature of websites and the need to meet NDA-expiry deadlines, well, ....demalion said:
BenSkywalker said:
Hmm...well, the quote you didn't include was talking about DX 9...the last time I ran the shader specific benchmarks I recall from the HardOCP article, they didn't require floating point. In any case, given 450MHz + bandwidth advantage versus 380 MHz, the results don't seem too surprising, as I'd expect reducing fp32 choke would allow better DX compliant optimization.
Considering this altogether, the shader results I'm thinking of from the HardOCP article have nothing to do with the misinformation I was discussing, were you talking about something other than the long list with the highlights indicating which result were highest?
BenSkywalker said:
Well, you're free to consider that equivalent for yourself, if you wish, but I'll simply point out to you that fp24 precision processing is hardly a secret we had to go digging for, and I think it is a pretty thin justification for the nature of Anand's commentary.
Okay, so first, impressions:
Pretty much what I had expected. I'd like to see nVidia's performance VS ATI's performance though, because there, nVidia's performance actually looks slightly better ( a little bit of tri ) so you might say it compensates ATI's AA advantage ( although that depends on tastes, I guess )
Also, I'm wondering if current benchmarks aren't a little too fillrate limited. Probably not most of them, but maybe we could see bigger performance boosts compared to the NV30 if we used less AF...
Now...
Sounds like I was either 100% wrong about the NV35's pipeline organization ( ) or that nVidia is doing a LOT of cheating, I mean, general cheating which would even apply in shadermark and all games.
I'm betting on a little bit of both
But WTF is the NV35 pipeline organization? And is it as impacted by register usage? We'd really need someone competent with cards to check all this...
2x FP & 2x FX units would be utopic. Or maybe each FP unit can only do one TEX operation ( that's fairly obvious, else it'd be a 4x4... )
Although nVidia would have done a fricking good job if they managed to fit all that in 130M transistors...
There GOTTA be a catch...
Or maybe my theory about decoupled PS/Tex is right, but that nVidia figured out a proprietary compiler technology to use FX only where it's perfectly invisible. Now that'd be EVYL! And nifty, too, because that'd be frciking great.
Uttar
Pretty much what I had expected. I'd like to see nVidia's performance VS ATI's performance though, because there, nVidia's performance actually looks slightly better ( a little bit of tri ) so you might say it compensates ATI's AA advantage ( although that depends on tastes, I guess )
Also, I'm wondering if current benchmarks aren't a little too fillrate limited. Probably not most of them, but maybe we could see bigger performance boosts compared to the NV30 if we used less AF...
Now...
Sounds like I was either 100% wrong about the NV35's pipeline organization (
) or that nVidia is doing a LOT of cheating, I mean, general cheating which would even apply in shadermark and all games.I'm betting on a little bit of both
But WTF is the NV35 pipeline organization? And is it as impacted by register usage? We'd really need someone competent with cards to check all this...
2x FP & 2x FX units would be utopic. Or maybe each FP unit can only do one TEX operation ( that's fairly obvious, else it'd be a 4x4... )
Although nVidia would have done a fricking good job if they managed to fit all that in 130M transistors...
There GOTTA be a catch...
Or maybe my theory about decoupled PS/Tex is right, but that nVidia figured out a proprietary compiler technology to use FX only where it's perfectly invisible. Now that'd be EVYL! And nifty, too, because that'd be frciking great.
Uttar
BenSkywalker
Regular
You were talking about misinformation here, and then later in your post you brought it up again. I was pointing out that {H} had benches that showed there was a doubling of shader performance in relation to the 5800U(although you have to go back to the prior review to cross reference it). The NV35 is showing ~2x shader performance compared to the NV30. Considering the trouble {H} goes through to level the playing field(Splinter Cell....) I doubt they are rigging the test up
I didn't say it was justified in the least, I was simply posting the misinformation that Anand was talking about. I made no commentary whatsoever as I am in no way supporting anything about the claims, you made mention that you didn't see an example of ATi's misinformation stated by Anand, I clearly remembered the quote so I posted it as is.
Joe DeFuria
Legend
antlers4 said:
Agreed. I've been saying that for quite a while now.
Doom3 is not designed to need FP precision to look great. What is needed to get "close to optimal" quality is to reduce the number of rendering passes. This is why you can run the R200 path on the R300, and not see much loss in image qulaity. This is why you probably WILL see a loss in image qaulity if you run the NV2x path on an NV3x card. It's not the floating point, it's the number of passes.
NV3x apparently runs intergers faster that floats, whereas the R3xx runs both about the same. This would explain why NV3X is significantly faster with the the NV3x path, vs. the ARB2 (floating point) path. It's also why the R300 runs both the R200 and ARB2 paths about the same speed.
The end result is likely this:
1) R300 will have the absolutely best image quality at its default setting. However, the differences will be in some things like "this dark corner over here...this weapon effect over there...this sky texture over here." Not some global huge image quality difference over the NV3x, because they both use similar numbers of passes ("single pass per light") for rendering. That being said, we all know review sites and fan-boys blow any image qulity difference out of proportion (see 32 bit vs. 22 bit rendering quality arguments with smoke trails, etc.). So even though Carmack says the quality differences are minimal (and to him, they probably are), it will be interesting to see how the internet community judges that.
2) NV3x can get that quality, by running ARB2, but will take a performance hit for it not to be worthwile.
Joe DeFuria said:
Joe - I was not really directing that at you. I apologize.
My comment was just in general. Some people are going to say the initial results are not correct. Or that they are correct. Or that Doom3 may run faster or not.
I threw out the conspiracy theories comment has people are already trying to find both positives and negatives on the card.
The problem here is that people are looking at the Doom3 results and thinking, "Wow! Great performance from the NV35. Wonder how NVidia managed it?"
Because NVidia have acquired the reputation of being less than honest in their marketing, people are automatically thinking that there is something funny going on with these results. I expect it will take a while before anyone accepts information coming out of NVidia at face value.
Because NVidia have acquired the reputation of being less than honest in their marketing, people are automatically thinking that there is something funny going on with these results. I expect it will take a while before anyone accepts information coming out of NVidia at face value.
Joe DeFuria
Legend
Bjorn said:
I'll just disagree with you on that one.
According to Nvidia (Tom's review), the ARB 2 path will run at the same speed as the NV30 path. But i think we will have to wait to see how that pans out
Yeah, nVidia driver code:
If PathRequset = NV30 then
UsePath = NV30
ElseIf PathRequest = ARB2 then
PathUse = NV30
End If
Well id don't know how Nvidia does the trick but there's no more IQ issues:
http://www.hardware.fr/articles/462/page6.html
http://www.hardware.fr/articles/462/page6.html