Has anyone heard if Nvidia are doing a 2-3W GPU?
It would outperform the best CPUs by 5-20x on parallelisable applications.
It could be used on battery applications (mobile), the same market that the Atom CPU is addressing.
This is a Killer app for Nvidia; there is nothing on the market in this segment. Nvidia probably already have the fastest Gflops per kW on the planet. If they get the technology right, then they could improve the Gflops/kW by factor 5.
The “Ion” platform is getting toward what you are talking about, but the power draw is about the same as Atom + Intel 945G motherboard, so 20-30W for the whole system.
At 2-3W, you’re talking system-on-a-chip, in which case the NVIDIA Tegra is close, but I don’t know if those devices (which are ARM11 + GPU on a chip) support CUDA in any form.
Sounds like NVIDIA are going in the low power direction with Tegra. I’m looking for a NVIDIA CUDA chip that can be the core of a COM Computer on Module drawing a total of 5-6W.