An industry at the crossroads
NVIDIA's VP of Investor Relations, Mike Hara, recently gave a pair of talks to investors in which he covered a variety of topics that are critical to the company's future: the discrete GPU market, AMD/ATI's R600 launch, Intel's forthcoming discrete GPU product, the high-performance computing market, and so on. Hara's comments on these issues are worth looking into, because they paint a picture of NVIDIA—and the graphics industry in general—at the first major crossroads since the erstwhile 3Dfx fired the opening shot in the 3D GPU wars. NVIDIA has picked one direction, and the other guys have picked two or three. The challenge for NVIDIA, then, lies in making the case that its single-minded focus on delivering the most GigaFLOPS per dollar will keep it ahead of the pack.
But before I dive into the substance of Hara's two very similar talks, there's a crucial bit of context that has to be laid out up-front. Hara has one mission here, and that's making the case to investors for NVIDIA's growth prospects. Because a publicly traded company's all-important stock price is much more indicative of the market's assessment of a company's revenue growth prospects than it is of profitability in any given quarter, everything that Hara says has to be viewed as part of the story that he's trying to tell about NVIDIA's plans to grow.
Product cycle and manufacturing process: next-gen GPUs on track for this fall
Hara began his talk at the recent JP Morgan analyst session by reiterating NVIDIA's commitment to a product cycle in which the high-end products come out each fall, and at the start of the next year, the midrange products follow, with the lower-end parts trailing later in the year. So at this point in the cycle, NVIDIA is addressing the lower end of the market while gearing up for another top-end refresh in the fall.
This fall will not only see the introduction of the next G80 derivative at the high end, but Hara also stated that the company will start to product its first 65nm GPUs. Moving its GPU line from 90nm to 65nm will bring NVIDIA to process parity with AMD/ATI (not that the 65nm feature size appears to have given the R600 any advantage in performance or power dissipation), and it will give the company the flexibility to increase either their performance or their profit margins, depending on how AMD/ATI's next-generation part fares.
Also on the topic of process technology, Hara laid out for an audience at Deutsch Bank Securities' 2007 Technology Conference NVIDIA's long-standing trick for getting good yields out of its GPU parts: reconfigurability. In a nutshell, reconfigurability lets NVIDIA fab most of its GPU products for the very high end and then test them to see if all of the pixel pipelines are functioning. If some pipes aren't functioning, then the company blows some fuses to reconfigure the device as a lower-end part with fewer pixel pipes and a lower clock speed.
GeForce 8800 Ultra and an 8800 GTS coming from the same 300mm wafer. One of the GPUs has two bad stream processing clusters, so after packaging and testing those two clusters are disabled, giving the part a total of 96 stream processors (compared to 128 on the G8800 Ultra).
Those of you who've followed IBM's Cell know that this is essentially what IBM is doing with the Cell processor's SPEs. Each chip is fabbed with eight SPEs, so IBM can determine which of the eight are fully functional and which ones need to be disabled. NVIDIA introduced this reconfigurability strategy with the GeForce 6 series, and they've been refining it ever since.
These types of reconfigurability tricks help companies increase yields and keep costs down by producing and packaging fewer duds.