The end is in sight for cheap GPU-based supercomputing, according to an International Science Grid This Week (iSGTW) opinion piece out this week. Author Greg Pfister argues that CUDA development has thus far been subsidized by the high volume sales of NVIDIA’s mass market low-end GPUs.
But sometime next year, we will see the arrival of chips that integrate the GPU and CPU on the same die. Intel’s “Sandy Bridge” processor chip and AMD’s Llano processor are both due out in mid-2011, and AMD’s CPU/GPU Fusion architecture is also in the works.
If the mass market consumer, using graphics mostly for games and entertainment purposes, can take advantage of these double-duty chips (as a referenced Anandtech article says they can) then where will be the market for NVIDIA’s low-end GPUs?
This means the end of the low-end graphics subsidy of high-performance GPGPUs like Nvidia’s CUDA. That subsidy is very significant, because the fixed costs of developing any chip family are very large; spreading them out over a high-volume low end makes a major difference, even if the high end has substantial revenue. So prices will rise, since GPGPUs will no longer have a huge price advantage over purpose-built HPC gear. How much will they rise? It’s very hard to say, but I have one somewhat wobbly data point saying that the difference will be substantial.
The ”wobbly data point” is arrived at by comparing a PS3 (mass market subsidized through volume and games) versus a custom built IBM HPC appliance, and extrapolating a 10 to 1 cost differential. Guess which one’s cheaper?
Even if the HPC market is growing as data suggests, Pfister notes that high-end GPU revenue is no match for the dollars generated by the demand for GPUs at the consumer level.
It may seem a stretch, but one way to still tap at least some of that mass market would be for NVIDIA to come up with their own integrated graphics processor. In fact, they already have, as noted in a recent HPCwire blog. NVIDIA’s Tegra line of processors, designed for mobile devices, has a heterogenous architecture including the low-power ARM processor and the GeForce GPU. Who’s to say NVIDIA isn’t planning another heterogenous processor using the CUDA-class chip with ARM CPUs? They could also choose to co-opt a higher-end CPU by teaming with another chipmaker. Or NVIDIA could even start designing its own x86 processor and then with that create an integrated graphics chip worthy of a gaming-class desktop or a Blue-Ray-capable notebook media machine. “NVIDIA-inside” anyone?