Cambridge University is turning to technology developed for computer gaming to help it solve some of the fundamental problems of science.
Cambridge University is using chips normally found in computer game consoles to drive cutting-edge research that could shed light on the origins of the universe.
The graphical processing units lie at the heart of Cambridge University's initiative to open up the power of supercomputers to mainstream academic research and businesses.
The commodity chips allow the Cambridge University’s supercomputing service to provide high performance computing power for a fraction of the cost of traditional supercomputer services, said Paul Calleja director of Cambridge's high performance computing service, in an interview with Computer Weekly.
“GPU computing offers power efficient computing and a factor of 10 improvement in performance per capita spend,” Calleja said.
Cambridge University's supercomputer centre provides scientists and local businesses with access to some of the country’s most powerful computing resources, running on Intel and GPU-based platforms.
Cambridge University's Intel-based supercomputer cluster was last week ranked as number 97 in the Top 500 list of the world’s most powerful computers. And it ranks as the most powerful cluster machine in the UK.
Downloadable resources on supercomputers:
Paul Calleja and his colleagues spotted the opportunity to create a community of GPU specialists two years ago.
“In Cambridge we had a good selection of academics that were looking at GPUs anyway, so we said, lets bring them together into a virtual group,” Calleja said.
The university approached graphics chip specialist Nvidia, winning funding to create a centre of excellence in Cuda, the programming language developed for its graphic chips.
Cambridge University teamed-up with Dell to build a GPU cluster, which features 128 Nvidia GPUs running on 32 Dell servers.
Bringing supercomputing to the masses
More articles on supercomputers
The GPU cluster has opened up supercomputing to a wider audience, said Calleja.
Many standard PCs contain Nvidia graphics processors. Researchers can go to the Nvidia web site, download a Cuda software development kit, and start programming.
“That is the strength of commoditisation and GPU computing, anyone can get started,” said Calleja.
Off-the-shelf open source programmes are readily available, covering everything from molecular dynamics and molecular modelling to materials modelling and computational fluid dynamics.
Staff at the Cambridge centre take the community codes, compile them for the GPU clusters and make them available to users, said Calleja.
“Probably 60% of our users are using some stock code. They are just consumers and it's not their code. They are not computer experts, they are scientists,” he said.
Creating evangelists to win consensus
However, it took effort to persuade academics in Cambridge of the merits of high-speed computing.
“We have had to put in quite a lot of effort to create a Cuda community. We say to people we will give you a GPU card if you act as an evangelist in your department, and we offer free computing time.”
Graphics chips work well for solving problems which require the same computational steps on multiple streams of data.
Turbine flow is good candidate for GPUs, said Calleja, because each blade requires its own set of calculations independent of adjacent blades.
And by opting to use GPU clusters, the Cambridge research group was able to reduce the cost of computing resources by a factor of 10.
“Instead of charging them £10,000 for a simulation, I only charge them £1,000. That has made a big difference to what they do,” said Calleja.
Molecular modelling for designing drugs and new materials, and gene sequencing are other strong areas of research.
The unit is working with a local hospital to provide real time analysis of CT body scans and MRI images.
The scanners take multiple pictures that need high power computing resources to knit together into a composite image.
“If that can be done it real time, it saves the patient a second hospital visit,” says Calleja. “It increases the throughput of the clinics dramatically. Effectively it doubles the throughput.”
At the other end of the scale, the service is working closely with scientists analysing data from the European Space Agency’s Planck satellite, which is mapping the background microwave radiation in the universe.
“We are trying to explore the very early origins of the universe, and that has lead to some very interesting discoveries. All of that work was done on a supercomputer,” said Calleja.
Cambridge is planning a major upgrade to its GPU cluster, with backing from industry.
The move will allow Cambridge to make GPU services available to network of high tech companies that have grown up around Cambridge University
“We are hiring a member of staff to work with the Cambridge network. We will be holding seminars to raise awareness, offer pump priming time on machines, and go through consultancy with them,” said Calleja.
Businesses in the area may not be looking for the answers to the origins of the universe, but Cambridge’s supercomputer should help them stay at the cutting edge.