NVIDIA Tegra Processors Blaze the Way for ARM in Supercomputing
As has become apparent to nearly everyone in the HPC community, life beyond petascale supercomputing will be power limited. Many efforts around the world are now underway to address this problem, both by commercial interests and researchers. One such effort that brings both into play is the Mont-Blanc research project at the Barcelona Supercomputing Center (BSC), which is looking to exploit ARM processors, GPUs, and other off-the-shelf technologies to produce radically energy-efficient supercomputers.
In this case, radically means using 15 to 30 times less energy than would be the case with current HPC technologies. The idea is to be able to build petascale, and eventually exascale supercomputers that would draw no more than twice the power of the top supercomputers today. (The world champ 10-petaflop K computer chews up 12MW running Linpack.) Specifically, the goal is develop an architecture that can scale to 50 petaflops on just 7MW of power in the 2014 timeframe, and 200 petaflops with 10MW by 2017.
The Mont-Blanc project was officially kicked off on October 14, and thanks to 14 million Euros in funding, is already in full swing. Last week, NVIDIA announced that BSC had built and deployed a prototype machine using the GPU maker’s ARM-based Tegra processors that have, up until now, been used only in mobile devices.
The power-sipping ARM is increasingly turning up in conversations around energy efficient HPC. At SC11 in Seattle last week, there were a couple of sessions along these lines, including a BoF on Energy Efficient High Performance Computing that featured the advantages of the ARM architecture for this line of work as well as a PGI exhibitor forum on some of the practical aspects of using ARM processors for high performance computing. There was also the recent news by ARM Ltd of its new 64-bit ARM design (ARMv8), which is intended to move the architecture into the server arena.
NVIDIA is already sold on ARM, and not just for the Tegra line. In January, the company revealed “Project Denver,” its plan to design processors that integrate NVIDIA-designed ARM CPUs and CUDA GPUs, with the idea of introducing them across their entire portfolio, including the high-end Tesla line. “We think that the momentum is clearly pointing in the direction of more and more ARM infiltration into the HPC space,” said Steve Scott, CTO of NVIDIA’s Tesla Unit.
The Mont-Blanc project is certainly an endorsement of this approach. The initial BSC prototype system is a 256-node cluster, with each node pairing a dual-core Tegra 2 with two independent ARM Cortex-A9 processors. The whole machine delivers a meager 512 gigaflops (peak) and an efficiency of about 300 megaflops/watt, which is on par with a current-generation x86-based cluster. The numbers here are somewhat meaningless though. The initial system is a proof of concept platform designed for researchers to begin development of the software stack and port some initial applications.
The second BSC prototype, scheduled to be built in the first half of 2012, will employ NVIDIA’s next-generation quad-core Tegra 3 chips hooked up to discrete NVIDIA GPUs, in this case, the GeForce 520MX (a GPU for laptops). This system is also 256 nodes, but will deliver on the order of 38 peak teraflops. Energy efficiency is estimated to be a much more impressive 7.5 gigaflops/watt, or more than three and a half times better than the industry-leading Blue Gene/Q supercomputer. In conjunction with this second prototype, NVIDIA will be releasing a new CUDA toolkit that will include ARM support.
The first two prototypes are BSC inventions. The project will subsequently develop its own more advanced prototype. According to Scott, that cluster will be 1,000 nodes, although the internal make-up is still not decided. Given the timeframe though (2013-2014), the system is likely to include NVIDIA processors using Project Denver technology, with the chip maker’s homegrown ARM implementation and much more performant GPUs.
By the end of the three-year project, the researchers intend to have a complete software stack, including an operating system, runtime libraries, scientific libraries, cluster management middleware, one or more file systems, and performance tools. They also hope to have 11 full-scale scientific applications running on the architecture, which encompass fluid dynamics, protein folding, weather modeling, quantum chromodynamics, and seismic simulations, among others.
Whether Mont-Blanc leads to any commercial HPC products remains to be seen. NVIDIA, for its part, is certainly happy to see this level interest and adoption of its ARM-GPU approach. “We see this as seeding the environment, where people can do software development and experimentation,” said Scott. “We think that it will grow into something larger down the road.”