The US Air Force Research Laboratory has put out a request for 2200 PS3s for supercomputing applications. Just in time for Christmas.
Scientists realized years ago that the PS3 Cell processor packs a lot of computer power at a very low cost.
The PS3’s Cell Broadband Engine processor, or Cell, is a heterogenous multiprocessor. Instead of identical cores - like Intel and AMD multi-core processors - the Cell consists of a 64 bit PowerPC core and 8 “synergistic co-processor elements” (SPEs).
Each SPE has 256 KB local store, a memory controller and a “synergistic processing unit” (SPU) with a Single Instruction, Multiple Data processing unit and 128 registers of 128 bits each. They’re connected by a bus with an internal bandwidth of more than 300 GB/s that transfers data between the SPEs.
The Air Force isn't going to play
Halo Uncharted 2 on them.
The new PS3s will be placed in a cluster environment with an existing cluster of 336 PS3s by connecting each of the units’ one gigabit Ethernet port to a common 24 port gigabit hub. Once the hardware configuration is implemented, software code will be developed in-house for cluster implementation utilizing a Linux-based operating software.
Only the PS3? The Feds have rules about sole-sourcing, so the Justification Review Document explains why only PS3s are acceptable:
With respect to cell processors, a single 1U server configured with two 3.2GHz cell processors can cost up to $8K while two Sony PS3s cost approximately $600. Though a single 3.2 GHz cell processor can deliver over 200 GFLOPS, whereas the Sony PS3 configuration delivers approximately 150 GFLOPS, the approximately tenfold cost difference per GFLOP makes the Sony PS3 the only viable technology for HPC applications.
What about Intel? The Air Force also looked at Intel:
. . . optimized performance on Xeon processors was explored through collaborative work with the Space Situational Awareness Institute at MHPCC. This work involved a detailed study of Xeon multithreading and SSE4 optimization on image processing intensive tasks. The subcluster headnodes are dual-quad Xeon servers. While these servers were found to be technically capable, they preclude a 500 TFlop/s system and their cost is more than an order of magnitude greater than the PS3 technology, thereby not fulfilling DHPI’s research needs.
Well, that settles it.
Update: A commenter reminded me that last year's World's Fastest Computer used a of lot of Cell CPUs. See more at PS3 chip powers world's fastest computer. End update.
What about GPUs? Graphics processing units (GPU) are the other hot parallel processing technology. And the Air Force looked at them too:
In the head-to-head competition between GPGPUs and Cell BEs, the cost of the host for the GPGPU was a significant factor. However, in this architecture, the subcluster headnodes are already provided, so the incremental GPGPU cost is attractive, even in comparison to PS3s. The GPGPUs have the additional advantage of not adding to the system size. . . . Our assessment is that the GPGPUs will 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.
In other words, GPUs are a good deal if you don't have to buy a computer to put them in. Which the Air Force didn't.
What do they do with them? I'm not sure the following answers the question, but it was the best I found:
The cell processor has shown large performance advantages for several applications, including Back Projection Synthetic Aperture Radar (SAR) Imager formation, High Definition Video image processing, and Neuromorphic Computing.
I Googled "neuromorphic computing" so you wouldn't have to. Wikipedia says
The term neuromorphic was coined by Carver Mead in the late 1980s to describe very-large-scale integration (VLSI) systems containing electronic analog circuits that mimic neuro-biological architectures present in the nervous system. In recent times the term neuromorphic has been used to describe both analog, digital or mixed-mode analog/digital VLSI systems that implement models of neural systems (for perception, motor control, or sensory processing) as well as software algorithms.
If I read that right a PS3 may be flying a real airplane someday. Uh-oh: Skynet!
The Storage Bits take The PS3 may be a close 3rd in the game console wars, but it is first in the hearts of high performance computing geeks. They've even published a do it yourself guide to build your own PS3 supercomputing cluster.
If you've been looking for an excuse to get a PS3, this could be it: "Honey, it's to improve my situational awareness - just like the Air Force!"
Comments welcome, of course. No, Sony has not sent me a review copy of the PS3. Maybe they're angry about my Blu-ray is dead posts? And Kudos to the AFRL for carefully managing the taxpayer's dollars. Also, credit to the Escapist where I first saw this story.