Nvidia Eyes Server Market With Future Tegra CPUs
Nvidia has stated that it prefers ARM to x86 for reasons of efficiency. ""The ARM instruction set is more power efficient than x86. That's why there are people looking to build ARM-based servers. That's why we like ARM in phones, because you get more performance per watt, more performance per square millimeter," Scott said.
The company many not be talking details yet, but it's not hard to picture what sort of Tegra-Tesla hybrid Nvidia might build. The efficiency and small die size advantages of ARM mean the company could conceivably design a Tesla GPU with an ARM CPU core baked into it. The OS scheduler and general compute functions would run on the ARM cores, while all the computational heavy lifting takes place on the GPU side of the equation.
One of the major advantages to this type of approach is scalability. Once it nails down the CPU side of the equation, Nvidia could ramp the size and capability of the associated GPU configuration to its heart content, with only relatively minor updates to the CPU. The 'Companion Core' in Kal-El could even be thought of as a variation on this theme. With Kal-El, NV embedded a small island of high-performing "G" transistors in the midst of a low power design. TSMC doesn't offer a design that allows the reverse, but NV is clearly experimenting with combining chip functionality and advanced manufacturing techniques.
The major hurdles facing Nvidia in the server room are related to ARM's limited performance relative to CPUs from AMD/Intel, and the fact that the vast majority of server software is written for x86 products. In this case, NV's chip is likely a considerable departure from other ARM products. While it remains based on the ARM instruction set, Nvidia has claimed its upcoming CPU will be 64-bit (all current ARM products, including the upcoming Cortex A-15, are 32-bit parts).
We expect Project Denver will eventually debut in several favors. A tablet/netbook-class part will consist of a chip configuration more analogous to modern-day Tegra 2, with a CPU and GPU on a single die. Eventually we'll likely see Tesla parts that include an ARM core, with said processor likely being on-package but off-die. Nvidia could conceivably integrate the ARM core directly into the GPU, but doing so would mean fabbing separate Tesla GPUs or burning out the CPU silicon when selling the chips in the enthusiast market.