Nvidia has announced partnerships with several high-performance computing (HPC) vendors to release servers featuring Tesla GPU-based accelerator boards alongside ARM processors - but not, oddly, those made by Nvidia itself.
The HPC market has been investigating a switch to the ARM architecture for some time, thanks largely to the rise of massively-parallel GPU-based accelerator boards like Nvidia's Tesla family. In supercomputers based on these boards, the CPUs in each node contribute a tiny fraction of the overall compute power; rather, they exist largely to keep the GPUs fed with data. A shift to slower but more energy-efficient ARM chips, it is hoped, will allow for a higher core count and thus even better throughout to the accelerators while also reducing overall energy envelopes.
The flaw in the plan is that GPGPU offload has largely been an x86 affair, with Nvidia also offering Power architecture support for its CUDA language. Recently, the company unveiled the first version of CUDA that supported GPGPU processing on ARM platforms - and has now unveiled partnerships with some big names in the HPC industry to take advantage of that support.
The new systems developed by Cirrascale, E4 Computer Engineering and Eurotech are based around Nvidia's Tesla K20 accelerator boards but not the company's Tegra family of ARM chips; instead, the CPUs are provided by Applied Micro, from the X-GeneARM64 family of 64-bit high-performance designs. Cirrascale will offer a 1U server with two accelerator boards, E4 a 3U dual-motherboard server with two K20 boards, and Eurotech a modular configuration based on its Brick Technology layout and featuring liquid cooling.
'Nvidia has built the industry's most comprehensive accelerated computing platform - including servers, software, development tools, processors, and related technologies - optimised for the HPC industry,' claimed Ian Buck, vice president of Accelerated Computing, at the formal unveiling during the International Supercomputing Conference this week. 'GPUs are the enabling technology that allow server vendors to build HPC-class systems around flexible ARM64 processors. The result is new, highly innovative computing solutions for HPC.'
Currently, Intel enjoys a majority share in the HPC CPU market, including providing both the Xeon CPUs and Xeon Phi co-processor boards for the world's fastest supercomputer Tianhe-2. Should Cambridge-based ARM's 64-bit designs prove capable, however, that could rapidly shift in a market where bleeding-edge adoption is always popular and the porting of applications to a new architecture an accepted reality.