Will Jocks strap on beefy ARMs to grapple rival racks next year?
Headlines to which the answer is no? Think again, popeye
Video When the University of Edinburgh decided to enter the gruelling cluster-building competition for students at this year's International Supercomputing Conference, they didn’t have to look far for hardware suppliers eager to feed their need for speed.
The server-wrangling undergraduates turned to UK-based businesses Viglen and Boston Ltd for the gear and smarts they’d need to compete at the cluster contest in Leipzig, Germany, last week.
I caught up with David Power, head of HPC at Boston, for a show floor discussion of what his company did to support the team (see the video below). The St Albans-based firm not only helped the students puzzle through the various options (Xeon vs. ARM, Phi vs. Tesla, etc.) but also jointly hosted the entire team for a three-day training session in London.
Team Edinburgh ended up using a Xeon-based system with NVIDIA K20 accelerators, driving it fast enough to land second place in the cluster competition's LINPACK competition with a score of 8.321 TFlop/s.
We may see a future Edinburgh team sporting ARM-based clusters, once the Brit-designed cores for servers become available. ARM tech is otherwise found in smartphones, hard disks and, well, just about everything handheld; the hype around high-end 64-bit versions of the cores is growing.
Boston is selling a super dense, low power, ARM server that utilises chips from Calxeda. This would certainly be the most energy-efficient option for student cluster builders, assuming that they could graft on accelerators to get the compute power they’d need to be competitive.
Check out the video to hear us discuss HPC application availability for ARM CPUs. From what I’ve learned from David and others, this shouldn’t be much of a problem. There are a fair number of scientific apps that have already been ported over to ARM, and the porting process looks to be relatively simple.
And bear in mind that's the whole point of these student cluster competitions: building machines against the clock that can crunch numbers the fastest within an electrical power limit. ®
Sponsored: Hyper-scale data management