Anandtech's "explanation" of NV3X's issues.

[I.S.T.]

http://www.anandtech.com/video/showdoc.html?i=2031 Or how I stopped worrying and learned to love just how biased Anandtech really is.

 

[akira888]

It did have one redeeming quality, a rather humorous malapropism in the title: "NV3X, a Moratorium." A "moratorium" refers to a period in which some activity is ceased, for example a moratorium on promotions or a moratorium on home foreclosures. What he meant most likely instead was "memorandum" which is a non-binding record of some past event - much more logical in the context of the article.

 

[nelg]

They could be referring to the time between a game release and its replacement shaders being introduced.

 

[akira888]

They could be referring to the time between a game release and its replacement shaders being introduced.

ROTFLMAO!

 

[DemoCoder]

I fail to see how Anandtech writing a post-mortem on the NV3x and it's problems constitutes a bias of Anandtech. This is exactly the kind of technical article I'd expect to see on B3D or GamaSutra.

 

[SiliconAbyss]

I thought the article was quite well done.

 

[kyleb]

not bad, dissapointing the way they overlook the fact that the nv3x was only as unimpressive becuase the r3xx are better. but then anandandtech has been overloking that in reviews for quite some time.

 

[Ylandro]

Except for the second part of their 'final words' I don't see much wrong with it.

I guess I missed one major point: Talking about the pipelines, they did indicate the difference to the R300. However, talking about the pixel shaders, there no mention at all about the R300... What kind of architecture is ATI using, and how does that compare to the changes NVidia made going from NV3x to NV40?

Anyone here who want to fill in that part?

 

[mjtdevries]

What is wrong with the article is that it does NOT explain what went wrong with the NV3x.

They explained that it is not easy to design a GPU, but they do not explain why Nvidia made the design decisions they did.
They make it sound as if Nvidia made very serious errors in creating their design, but Anand still claims at the end that the design was not flawed.

It's not enough to explain that a 4x2 is at a disadvantage when using single textures. You have to explain why Nvidia decided to use 4x2 even though their engineers must have known about that disadvantage.
You have to explain why they tought 4x2 would be an advantage. (otherwise they would not have choosen it)

When you have to draw a conclusion purely from the information in the article, the only conclusion can be that the Nvidia engineers are utterly stupid in making those silly design decisions. And that the ATI engineers made the logical decisions, that everybody can easily understand are the best solution.

I'm not a fan of Nvidia (I cannot be a fan of a company that thinks it's customers are stupid and can be cheated). But I do not believe it's engineers are stupid.
They have made fine products in the past, and seem to have made a solid product in the form of the NV40 too. They must have had a valid reason to design the NV30 the way they did, and I don't see any reason in Anand's article.

Personally I think the reason is that they were not asked to design a DX9 GPU, but were asked to design a DX8 GPU with DX9 just being a checkbox feature.

Seen in that light, the design of the NV30 was not flawed. It did exactly what it was designed to do, and did that well.
In that light it was the vision of Nvidia that was flawed.

 

[nelg]

I fail to see how Anandtech writing a post-mortem on the NV3x and it's problems constitutes a bias of Anandtech. This is exactly the kind of technical article I'd expect to see on B3D or GamaSutra.

This is exactly the kind of technical article I'd expect to see on B3D, except it would be a preview not a post-mortem. It just seems a day late and a dollar short.

 

[akira888]

Despite the fact that the achilles' heel of the NV3X architecture (expecially NV31 + NV34) was the relative lack of floating point fragment shading units the author failed to directly compare the shading/texturing abilities of the NV3X against R3XX. Which at least to me seems be the first thing any post-mortem of NV3X should have somewhere.

 

[DemoCoder]

Writing previews and writing port-mortems are different. You often don't have enough data or information at the preview stage to make the same conclusions you can in post mortems. B3D couldn't write about all the issues with the NV3x because there weren't enough tools at the time to reverse engineer what's going on and IHVs are less likely to give out proprietary information on the hardware implementation at the preview stage. Things like register limitations came out much later.

Secondly, post-mortems don't usually make comparisons against other products. Go read gamasutra's post-mortems and see if they make comparisons. Post-mortems are a survey of "what was done right, what was done wrong, what you wished you did differently, and why"

I don't know why you guys expect an NV3x post-mortem to even mention the R3xx. B3D's reviews don't do cross-product comparisons either.

The worst thing you can say about the NV3x post-mortem is that it should have been written by NVidia employees, since post-mortems are usually written by the people involved.

 

[Reverend]

You often don't have enough data or information at the preview stage to make the same conclusions you can in post mortems.

Add in "not enough time before meeting worldwide publication date".

I thought the article was a good read but not exactly a good attempt at what it proposed to do (i.e. study the NV3x's flaws).

 

[nelg]

I agree with what you are saying DC but the issue I have with the article is that Anandtech has done very little until now to get under the skin of the Nv30. If Beyond3D was to do such an article it would be just a summary of what already has been investigated and discussed. If you were to read Anandtech exclusively, this would have been news. Insofar as having it make a comparison to the R3xx, I have no such expectation.

 

[mjtdevries]

I agree that this information might not have been available at the time when previews of the NV30 were made. (Especially since we didn't know for sure at that time whether it was a 4x2 or 8x1 design)

But lots of sites have from the very first moment spent time to determine that. So this information could have been included in a review. Especially since a lot of people would have wanted to know AT THAT TIME why the NV30 performance was lower then expected. Now it is a bit too late.


About the comparison with the R300. It's not strictly necessary, but it is logical to include a comparison. The reason for that is very simple:
What went wrong with the NV30 is that ATI had made a card that was better.

I'm not even sure anything went wrong in the design of the NV30.
I think that it performed exactly as Nvidia had planned. But they had not planned that ATI would have a card that could easily match it in DX8 performance, and ALSO had very strong DX9 performance. THAT is what went wrong with the card.

As I said before, I think Nvidia designed the NV30 to be a DX8 card with DX9 as checkbox features. That works when your competition doesn't have any DX9 features. IMO they have used this design strategy several times in the past with success. It just didn't work this time.

 

[Arun]

Errr I'd love an explanation of this line if someone got any clue about it:

Further improvements in NV40 were made to the help eliminate hidden pixels earlier in the pipeline at the vertex shaders

Either NVIDIA has some kind of Area-51 tech here, or Anandtech just doesn't know WTF they're talking about. I'd bet on Option 2, but I'm still wondering where that idea comes from hehe

Uttar

 

[Joe DeFuria]

My main gripe is similar to others...this shouldn't have been a "post mortem". It should've been 6 months to a year ago.

They use NV40 architecture to "compare and contrast" to the NV3x architecture...when there should have been an article doing the same a year ago with NV3x vs. R3xx.

Especially since a lot of people would have wanted to know AT THAT TIME why the NV30 performance was lower then expected.

That was easily brushed aside in the past for two reasons:

1) Evanglization that the NV3x core was so "complex and advanced" that "new drivers" will be needed to really get at the performance potential.

2) Dismissal of benchmarks *cough* 3DMark *cough* that actually SHOW that Nv30 performance was lower than expected.

It's no surprise that Anand would have such an article up now....after nVidia has launched a part that replaces the NV3x...

 

[WaltC]

http://www.anandtech.com/video/showdoc.html?i=2031 Or how I stopped worrying and learned to love just how biased Anandtech really is.

I just read it and was going to do a post here myself about it, but you beat me to it... I agree, and I thought the whole thing was little more than an apology for nV3x that might've been penned (well, at least coached from the sidelines) by nVidia itself. The whole problem with it is that the article seems to depend on nV40 as a reference for "what was wrong with nV3x" when the R3x0 architecture has been serving the purpose of illustrating what was wrong with nV3x for the past 18 months...

This article would have been timely and somewhat informative 8-12 months ago, using R3x0 as a reference (nV40 not required.) As it is, though, it's just an apology the purpose of which is to try and promote nV40 because it is "so much better" than nV3x. The truth is that "what was wrong" with nV3x is that it fell far short of R3x0, and what is "right" about nV40 is that its design is much closer to that of R3x0 than nV3x ever was.

I think a good sidebar option for this article might be: "What took us so long at AnandTech to figure out what was wrong with nV3x?"

The article did, finally, put some facts into writing that AT should have written about long ago. Notable were these:

Behind NV3x is a 4x2 pixel pipe (though there was some confusion over this we will get to later). ...

...rather than coloring a pixel, a z or stencil operation can be performed in the color unit. This allows NV3x to perform 8 z or stencil ops per clock and NV40 to perform 32 z or stencil ops per clock. NVIDIA has started to call this "8x0" and "32x0", respectively, as no new pixels are drawn.

(Empahsis mine.) Finally, the fact that "ops" per clock in no way resemble pixels per clock, is at last sinking into the mainstream hardware press enough to actually see its way to print... Only took them--let's see--about 18 months to understand the very simple premise. I guess we can rejoice over small miracles whenever they occur.

So why did so many people trip over their shoelaces over something so small as understanding the differences between "ops" and "pixels," differences which are substantial, fundamental, and obvious? I think the following quote from the article explains it:

Of course, there is more to graphics performance than how many pixel pipes are under the hood.

Of course there is, and no one has ever said differently. The problem was that some interpreted that since there's "more to graphics performance in a gpu than the number of pixel pipes," it means, somehow, that the number of pixel pipes in a given gpu isn't important or germane to a discussion of that gpu's performance, or that the number of pixel pipes in a given gpu were flexible to the degree that pixel pipes and the number of ops per clock were freely interchangeable in a discussion if not indistinguishable in reality. Which of course is sheer nonsense...

This "confusion," in my opinion, originated after nVidia misrepresented the pixel pipeline organization of nV30 as "8x1" when it was later revealed that nV30 was clearly 4x2 (despite nVidia's official representation of the 8x1 spec.) nVidia apologists came out of the woodwork to "explain it" away under a blanket of confusion which did little more for me except to underscore that they themselves didn't know the difference between a pixel and an "op." To me, it was analogous to not knowing the difference in the human body between a leg and a toe...

Prior to those events, I do not recall any such general confusion existing as to the fact that all gpus have a fixed number of pixel pipes in their physical architectures, and that knowing that number was a decent baseline from which to begin a meaningful investigation into the target gpu's overall performance profile. It's just a baseline, but it is a fundamental baseline to have if enlightenment is expected. Without such unambiguous knowledge of baseline specifications, confusion will reign.

Where the article here well and truly falls short of the mark is in the fact that the blame for the "confusion" has to go to nVidia, since nVidia was content to allow such misrepresentations and misunderstandings to fester and grow so long as nVidia perceived some direct PR advantage in doing so. That was obviously the company's motivation behind dishonestly representing the 8x1 organization in the first place (since R300 was indeed 8x1 and nVidia was cognizant of the difference): nVidia was not in the least confused about why it would be better to state an 8x1 organization for nV30 (even if it was in reality 4x2.) Had nVidia not thought such a statement, even if dishonest, would help it in some fashion, it would never have made the bogus 8x1 claim for nV30 in the first place. No one who designed nV3x is in the slightest "confused" about the fundamental differences between an op and a pixel, I'm quite sure.

I thought this next comment was particularly insightful, and was glad to see this make its way into print as well:

They needed the performance leap, and now they will be in a tough position when it comes to making money on this chip. Yields will be lower than NV3x, but retail prices are not going to move beyond the $500 mark.

(Emphasis mine.) If the yield picture is indeed worse than it was for nV30 (which was cancelled) or nV35/38 (which didn't see the light of day in any sort of "quantity" until late in the 3rd calendar quarter of '03 and beyond), then the picture for nV40 is extremely bleak. Time will tell about this, though, and if nVidia can successfully yield mass-market quantities of nV40, a gpu manufactured on the same .13 process as nV35/8, but with nearly *double* the transistor count, my hat will be off to them as this will be quite an achievement, indeed. Because, just as the question, "If a tree falls in the forest where no one can hear it, does it make a sound?" is an interesting one in some respects, so is this question: "Will nV40 actually matter if nVidia is unable to ship it in meaningful quantities?"

In 30-90 days we should have an answer to the vexing philosophical question: "How many nV40s can you fit on the head of a pin?" The answer must be "all of them" or "none"... If the answer is "all of them" it will mean yields are too poor for the nV40 to sell in meaningful quantities; if it is "none" it will mean yields are OK and even one nV40 is too large to fit...Heh... OK, enough philosophy...

Edit: typos

 

[digitalwanderer]

Where the article here well and truly falls short of the mark is in the fact that the blame for the "confusion" has to go to nVidia, since nVidia was content to allow such misrepresentations and misunderstandings to fester and grow so long as nVidia perceived some direct PR advantage in doing so. That was obviously the company's motivation behind dishonestly representing the 8x1 organization in the first place (since R300 was indeed 8x1 and nVidia was congnizant of the difference)

Yup and thanks Walt, I couldn't put my finger on why I didn't like the AT piece. They seemed to slant it that nVidia was the victim of a serious of unfortunate luck and some decisions which in hindsight don't seem like they might have been the most prudent....they seemed to have missed nVidia's rather active role in perpetrating and carrying on all the misperceptions and misinformation.

If the yield picture is indeed worse than it was for nV30 (which was cancelled) or nV35/38 (which didn't see the light of day in any sort of "quantity" until late in the 3rd calendar quarter of '03 and beyond), then the picture for nV40 is extremely bleak. Time will tell about this, though, and if nVidia can successfully yield mass-market quantities of nV40, a gpu manufactured on the same .13 process as nV35/8, but with nearly *double* the transistor count, my hat will be off to them as this will be quite an achievement, indeed.

You don't have to wait Walt, I can answer that one right now....

....JULY!

 

[StealthHawk]

This article would have been timely and somewhat informative 8-12 months ago, using R3x0 as a reference (nV40 not required.) As it is, though, it's just an apology the purpose of which is to try and promote nV40 because it is "so much better" than nV3x. The truth is that "what was wrong" with nV3x is that it fell far short of R3x0, and what is "right" about nV40 is that its design is much closer to that of R3x0 than nV3x ever was.

Exactly what I thought of the timing.