When Sony issued a recent PlayStation 3 update removing the device's ability to install alternate operating systems like Linux, it did so to protect copyrighted content—but several research projects suffered collateral damage.

The Air Force is one example. The Air Force Research Laboratory in Rome, New York picked up 336 PS3 systems in 2009 and built itself a 53 teraFLOP processing cluster. Once completed as a proof of concept, Air Force researchers then scaled up by a factor of six and went in search of 2,200 more consoles (later scaled back to 1,700). The $663,000 contract was awarded on January 6, 2010, to a small company called Fixstars that could provide 1,700 160GB PS3 systems to the government.



Another grotesque waste of taxpayer dollars? Exactly the opposite, according to research lab staff. Off-the-shelf PS3s could take advantage of Sony's hardware subsidy to get powerful Cell processors more cheaply than via any other solution.

"The Advanced Computing Architectures team at the Information Directorate considered several alternatives to arrive at the configuration of the proposed system, including the Sony BCU-100, IBM Blade Q22, and IBM PowerXCell 8i CAB accelerators cards," said the Air Force last year. "In particular, the performance capabilities of the Cell Broadband engine were examined in considerable detail on each of the algorithms."

The team also looked into using dual-quad-core Xeon servers for its cluster, going so far as to do a "detailed study of Xeon multithreading and SSE4 optimization on image processing intensive tasks." The hardware worked well, and it eventually came to serve as subcluster headnodes that sit between the PS3 cluster itself and the control terminals.

But building the entire cluster out of Xeons would cost "more than an order of magnitude greater than the PS3 technology." The team also looked into advanced GPGPUs but found that they worked best to "accelerate a subset of our algorithms, particularly the frontend processing and backend visualization, but lag the PS3 in the bulk of the calculations where processes need to intercommunicate and share memory beyond what is supported efficiently by the GPGPUs."