New supercomputer is a rack of PlayStations

The Age - February 26, 2008
Louisa Hearn


When the PlayStation3 was released in November 2006, Gaurav Khanna's wife braved long queues so he could be one of the first people in the US to get his hands on the gaming console.

But the astrophysicist was not itching to burn some rubber in Gran Turismo or shoot hoops in NBA 07. Instead he wanted to build his own supercomputer.

Mr Khanna now owns 16 PS3s, which spend their days simulating the activities of very large black holes in the universe for the physics department at the University of Massachusetts.

Hooked together in a single cluster, the PS3 consoles provide his department with the same amount of computing power as a 400-node supercomputer.

"The challenge these days with supercomputing facilities is that there is a lot of demand for them. So even if I submitted a job that would be expected to take about an hour, it could actually take two days to get started because the queues are so long.

"The PS3 cluster is all mine and was very low cost to set up, which makes it really attractive," he says.

What makes the gaming console vastly superior to high-end computers for complex research algorithms, Mr Khanna says, is the Cell chip built by IBM to facilitate high-end gaming functions on the latest generation of consoles.

In addition the PS3 was built with an open hardware architecture, which can run the Linux operating system.

Based on Einstein's theory of relativity, Mr Khanna's research on black holes is purely theoretical. In order to run his simulation data on the console he has to reprogram it so the algorithms will work on the new architecture.

"Linux can turn any system into a general purpose computer but for it to do work for me I have to run my own code on it for astrophysics applications. The hard part of the job was to make sure my own calculations could run fast on the platform, which meant I had to optimise the written code so it could utilise the new features of the system.

"I am not a Linux person - I am a Mac person - but I was able to follow instructions online," he says.

His next challenge will be to turn his data into graphical simulations using the high end graphics engine included in the PS3.

"We haven't done that yet but it would be very neat to actually see the simulation while it is going on," he says.

Although Mr Khanna was one of the first scientists to optimise the PS3 for his own research work, Tod Martinez, Professor of Chemistry at the University of Illinois has been tinkering with games consoles ever since his son's original PlayStation malfunctioned.

He says that while researching whether or not to buy a new PS2 for his son he also began to explore the possibilities of using gaming consoles for scientific research.

"The two main things we do are rotations and translations of objects. We also need to get a lot of pixels onto the screen, which means we need big channels to move lots of data. It was pretty clear that modern games consoles mapped really well onto theoretical chemical calculations," he says.

Once the PS2 was released he bought one unit for his son and a few units for himself and made some rudimentary attempts to program it. "But back then the architecture was proprietary and trying to convince the machine to run non-Sony programmed games was difficult."

That situation improved rapidly when Sony released a DVD that would allow users to run Linux on the console. "Only 1000 DVDs were released but they really opened up the architecture considerably," he says.

Since that time Mr Martinez has refined his computing resources considerably and he now runs a cluster of eight cut-down consoles from IBM based on the same Cell chip technology used in the PS3.

He has also expanded his use of gaming technology into graphics cards with the help of a new programming framework developed by graphics hardware specialist Nvidia.

Mr Martinez's field of research is examining how molecules behave when you shine light on them - which has wide ramifications for the fields of agriculture, solar energy and the study of human vision.

"We have done tests of algorithms and Nvidia cards are four to five times faster than the Cell chip, which is 20 times faster than an ordinary high end computer," he says.

Because both technologies can be classified as "stream processers" they are highly suited to moving massive volumes of data - unlike the general purpose processing for ordinary computing.

"Some people think it's just about having a faster computer. They don't realise how big a change it is to do computing at your desk after accessing a computer in a room somewhere where you have to wait around for results.

"Of course it does cost less, but what needs to be recognised is that it also changes the way people think about problems when they are given a hundred times more computer power."

Using the example of a black box, Mr Martinez explains that instead of asking basic questions about how it works, you can just start tinkering around with it.

"So rather than taking the thing apart you just start moving all the knobs about to see what happens when you change something - just as you might in real life."