We would like to take this opportunity to clarify the Mont-Blanc project vision in order to complement the comments made in the interesting interview with the leading HPC experts Thomas Sterling and Satoshi Matsuoka in Primeur Weekly Magazine last week ( 2014 Another year on the road to Exascale – An Interview with Satoshi Matsuoka and Thomas Sterling – Part I ). The Mont-Blanc project indeed advocates for the use of ARM IP blocks in future Exascale-class supercomputers, and as detailed in project related publications such as our SC’13 paper, we quickly reached the conclusion that ARM processors are not intrinsically more energy-efficient than their Intel and AMD counterparts. In this sense we agree with the comments made in the interview.
Mont-Blanc was the first project to suggest using ARM chips in supercomputers and has taken the first steps to demonstrating their viability in this context. In the very near future we expect to see commercial supercomputers based on ARM chips. In fact many leading vendors are already working in this direction.
We advocate for the SoC integration leading to higher energy efficiency. As mentioned in thePrimeurinterview, it is the freedom to integrate the right set of components and customize design parameters to the target workload that opens opportunities, not the use of a “low power” processor IP. This is reflected in the contents of Alex Ramirez’s presentation in ISC’14 (Leipzig): One Size Fits All vs. System-on-Chip Integration for HPC.
The Mont-Blanc project has made significant contributions in this direction, mainly enabling the use of ARM processors and GPUs in HPC, porting applications, libraries and tools to build an ARM cluster that is backwards compatible with current HPC clusters. We like to believe that, in a small way, we are doing something comparable to what was done by Thomas Sterling, in one of his great contributions to the field, when he demonstrated that clusters could be used for HPC.
Mont-Blanc is limited to a conservative approach by one main factor: funding. Given this limitation, instead of building a custom ARM-based HPC SoC with an integrated network interface, GPU, non-volatile memory controllers, and integrated 3D memories, the project has had to use commodity SoCs from the mobile market, which offer some of the desired features and still produce a competitive performance / (energy * cost) solution. This is something we have already demonstrated in the Montblanc project and the previous Pedraforca and Tibidabo prototypes.
The proposed continuation of the Mont-Blanc project adapts to these limitations by focussing on software and leaving hardware development to those who have a real commercial commitment.
BSC has defended for some time that Europe is capable of designing competitive HPC systems. However, in order to do this, commitment from the EC, the European HPC industry and European users is needed. If this commitment is forthcoming, we will be able to produce ARM-based chips which would allow competitive systems of all levels to be built: for mobiles, servers and HPC. This is the challenge we are facing, and we believe that the Mont-Blanc project can play a key role in meeting it.
Mateo Valero, Director, Barcelona Supercomputing Center, Mont Blanc project