ARM+GPU = dream super?
NVIDIA: 'Yup, we're on it'
SC11 We've seen a lot of ARM server activity in recent weeks, with ARM chip upstart Calxeda announcing its 5-watt EnergyCore ARM chip (breakdown here) and a partnership with x86 giant HP (outlined here). Also the ARM architecture is getting a necessary but painfully slow 64-bit makeover, slated to roll out in 2014.
ARM processors use much less power than even the stingiest x86 chip and for this reason many pundits tout them as the answer for massive web serving infrastructures. But ARM simply doesn't get it done when it comes torigorous enterprise and HPC (high performance computing) processing chores. The 32-bit versions can't address enough memory, and there the ARM chip and instruction set lacks the oomph to compete with mainstream x86 processors on server-like chores.
But combining ARM with the massive number-crunching power of GPUs might give HPC and analytic processing types the best of both worlds: very high energy efficiency and density without performance compromises.
An approach like this is possibly the best route to exascale, where the biggest hurdle is that simply powering up an exascale system will take the juice of a small city - and a fair amount of its real estate too.
Climb every mountain
At SC11 on Monday, NVIDIA announces major progress along these lines. First, the company will talk about Europe's Mont-Blanc project, the goal of which is to develop a European exascale approach using ultra energy-efficient, off-the-shelf components to build supercomputers capable of competing with Top500 leaders.
Mont-Blanc will rely heavily on NVIDIA's Tegra ARM processor to act as traffic cop for GeForce GPU accelerators. The first prototype will have 256 ARM processors fronting an unspecified number of NVIDIA GPUs. There is a discussion of Mont-Blanc plans here.
Help is on the way if you want to investigate the possibilities inherent in combining ARM and GPUs, but don't have the check book of a major international computing consortium behind you.
Next year will bring a CUDA for ARM development kit that includes a board with a quad-core NVIDIA Tegra 3 ARM processor and a CUDA-enabled GPU, along with Gigabit Ethernet, SATA and USB ports. Pricing and other details aren't yet available.
The combination of ARM and GPUs (or other specialized accelerators) poses the strongest challenge yet to the x86 hegemony in scientific and technical computing. Admittedly, there is not much of an ARM HPC software ecosystem, but HPC users are typically quite willing to roll their own when it comes to new technology. They're happy to build and optimize their own software if the end result is higher performance at less cost.
If the combo proves successful in HPC, then it's a matter of time before it becomes more fully commercialized and starts popping up in enterprise data centers. ®