When you factor in the significantly lower cost of off-the-shelf versus custom solutions, the Air Force’s decision seems to make more sense, particularly now that calls for fiscal restraint are getting louder.

Finally, government projects require significant overhead to cover installation, maintenance, and (as mentioned previously) customization. That should account for most of the cost differential that you describe.

Besides that, the PS3 Cell is crippled with only 6 active SPEs and 256MB of RAM. Plus, it’s the original single-precision floating point version without the pipelined double-precision engine, so any high precision numerical computing tasks will suffer a rough 10x performance penalty.

Also, forget about using the GPU, as that’s locked out on the basic Linux install. As for connectivity, the best it can do is gigabit ethernet. The PS3 sounds attractive due to it’s price and off-the-shelf nature, but honestly, I feel as though there are better solutions.

IBM created a blade, the QS22 (http://www-03.ibm.com/systems/bladecenter/hardware/servers/qs22/) for computing with the Cell. It has 2 full (8 SPE) Cells with up to 32GB of RAM and an Infiniband interconnect if one desires. Granted, it’s much more expensive, but it’s purpose built to crunch numbers.

Finally, at retail, 2000 PS3s should only cost roughly $600,000. What are they doing with the other $1.4MM?

Interestingly, Compaq also did not make any money selling their supercomputers to national laboratories. Instead, they saw it as a PR opportunity that would help them sell high end servers to their corporate customers. I am not sure that Sony can claim the same type of benefit.

One of the more interesting projects that will be conducted at the laboratory is neuromorphic computing. These are early attempts to create intelligence that truly mimics the complexity of the human brain.

In “Neuromorphic Computing: Teaching New Brains Old Tricks,” Brad Minnery observes that “Deep Blue may be able to mop the board with most humans in a game of chess, but pit IBM’s vaunted supercomputer against your average canine in a game of ‘fetch my slippers,’ and [the dog] wins every time.”

On the other hand, neuromorphic computing could one day provide “adaptive computer interfaces, intelligent tutoring software, personal devices that sense a user’s mood and social context, prosthetic devices for the disabled that integrate directly with the nervous system, and even intelligent household robots that do all the little things—like retrieve a pair of slippers.”

Minnery also claims that this research could have implications for the video game industry.

The same neuromorphic learning algorithms that may enable a military ground robot to learn local terrain features could also endow commercial video game characters with the ability to learn and adapt to the tactics of human players.

Of course, how much value that will have in real life remains to be seen. Currently, gamers can obtain the same results by going online and competing against any number of human opponents.