6. The NV30 debacle

3Dfx Rampage was to be called Spectre 5600/5800. It’s sad NVIDIA decided to name “FX 5800” after such a weak product. I guess we didn’t say “RIP, 3DFX” enough. Or maybe they just didn’t listen enough. Like the FX 5800, Spectre 5800 would have been a nice product if it had shipped on time. Unfortunately, neither 5800 shipped on time, and today’s 5800 can’t compare with ATI’s 2003 product line.

The saga began in late 2000, slightly before NVIDIA bought off 3DFX intellectual property. The NV30 had been under development for a few months, yet the 3DFX acquisition resulted in a redesign of the NV30, with estimations of only a slight delay, to SIGGRAPH 2002. The original plans seem to have been to launch the NV30 approximately at the time the GF4 was launched (which did have some 3DFX influence since its 2D engine very closely resembles Rampage’s).

After all, NVIDIA seemed omnipotent at the time. A slight delay couldn’t kill them. Unfortunately, it wasn’t a slight delay for long. Nothing went right after this decision. The first problem seems to have been the time it took to decide the final design, getting both NVIDIA and 3DFX employees productive. The second was a general underestimation of all related risks, uncertainties, and potential implementation delays. The final problem, which is really the biggest one IMO, is a massively exaggerated optimism about the specs, the obviously awful paper-to-silicon transition, as well as about yields.

The original designs were completely out of this world: 500/1000 (yes, even the original designs wanted 1000Mhz 128-bit memory bus GDDR2, so that’s clearly a big design mistake), Programmable Primitive Processor ( PPP ) support as in the NV40/R400/R500, and multichip (the exact signification of multichip in that context remaining mysterious at best).

There were so many redesigns, no one’s would ever accept to count them. Probably as many ones as the mythic Rampage – pretty big changes, too. The PPP went poof, and the design went from 6x2/12x0 to the current 4x2/8x0. Heck, if we imagined the NV35 as a 6x2/12x0, NVIDIA would significantly beat the Radeons in just about every benchmark but Pixel Shading intensive ones – if 6x2/12x0 means what I think it means, at least. But it obviously didn’t happen that way.

But if they did so darn bad, where did the $347M go? One possible explanation is that their “free caviar and champagne” parties were just a TAD too extravagant. But someone, who I hope, will excuse me for quoting him here, gives a much more sensible explanation:

nV had seven NV30 tapeouts (each costs around 10 mil), payed extra 20 mil for speeding up wafer production and in the end, only 4th or 5th tapeout chip came back in working state....it was pure panic.

Each wafer can have 126 or 127 (not sure) NV30 chips […] They went into mass production with 7-10 chips per wafer.

Each NV30 chip cost nV around 60 bucks. Instead of 10-20.

[Correction, 11/9: The $10M/tape-out cost seems significantly too high, it should be more like $1M – so either that was a typo from the source, or it included a lot of other barely related costs. Also, the 7-10 chips per wafer number seems to be for one of the first working tape-outs where NVIDIA absolutely needed chips to show to the world – the $60 per chip cost would then be the lowest NVIDIA managed to put costs at, after several respins. Also, there most likely weren’t 7 tape-outs, maybe rather 4 tape-outs and 3 respins. - Uttar]

Marketing & PR played a pretty big role in intensifying the problem, by increasing clock speeds and indirectly forcing Flow FX to compensate. The massive hyping of the NV30 was once again caused by misinformation and incompetence. It seems to me, based on the lack of information and lack of understanding by non-engineering departments, they placed their bets on optimism and hyped NV30 like a killer. Unfortunately, the only thing killed was NVIDIA’s reputation.

I won’t go in the exact reasons, the NV3x currently fails to deliver acceptable Pixel Shading performance. When is NVIDIA going to fess up that not only the NV30, but also the NV31/34 are bad GPUs, and except for workstation work, don’t offer adequate performance for the money… Never, I suppose, since they're money-making parts, unlike the NV30... The NV35/36 are much better IMO, although against ATI’s offerings, of course they don’t compete well.

With such spectactular string of failures, what could NVIDIA originally have had in mind? First of all, full VS3.0. compliance, (ab)using the lookup units. At the same time, the PS was ( and is ) capable of using the COS/SIN tables, which are primarily but not exclusively useful to the VS.

The professional market justified these decisions, as the NV30 design seems to have been based on the workstations’ needs, as the Russian DAWN to DUSK conference highlighted:

The NV30 and the NV30GL don't differ much; moreover, first of all the company developed a professional version and then adapted it for use in game accelerators.

I’m uncertain, but I’ll speculate it’s because Marketing and/or Management wanted to be able to claim “Cinematic Computing” for consumer gamers, despite being a feature advantage which could only benefit a small part of the workstation market.

Another big mistake was to focus on making the “historically slow” go faster, rather than to improve overall performance making COS/SIN significantly faster than on the R300 seems rather futile when you realize the basic shading functions of the chip haven’t improved much since the NV25 on a clock-for-clock basis.

The decision to implement FX12 units that are below the DX9 PS2.0 absolute minimum of FP16 might have been caused by NVIDIA delaying the DX9 input agreement, as mentioned even by the recent ATI presentation leak, in the part regarding Cg’s reasons to exist.

I thus conclude three main historical reasons for which the NV30 design failed:

• Massively exaggerated optimism and prentension from all NVIDIA departments.
• Marketing decisions forced Engineering to focus on features, and not performance.
• Bad design choices, sometimes illogical, sometimes against standards, and sometimes just making the best of a bad situation.

I could go further, explaining my opinions on legitimate driver improvement opportunities, but that’s not my goal here. If I’m writing about these problems, it’s not to publish a complete list of specific bad decisions or products, but instead it’s because…