According to http://www-132.ibm.com/webapp/wcs/stores/servlet/CategoryDisplay?categoryId=4611686018425192155&storeId=1&catalogId=-840&langId=-1 http://www-132.

According to http://www-132.ibm.com/webapp/wcs/stores/servlet/CategoryDisplay?categoryId=4611686018425192155&storeId=1&catalogId=-840&langId=-1 http://www-132.ibm.com/webapp/wcs/stores/servlet/CategoryDis... a blade with two cell chips costs $10k. Two PS3s come in quite a bit cheaper than that. Okay, so they don't have as fast an interconnect and as much RAM, but it depends on your problem whether that matters at all. Modern graphics cards are also useful for number-crunching, but it's pretty awkward to develop general-purpose code for them. Not that the Cell is all that easy to code for with its SPE-local SRAM and explicit DMAs for memory access. As far as I can tell, though, the Cell development tools are more mature, and it wins in terms of FLOPS-per-watt. I suspect the real win for the Cell is that its interconnect latency is tiny compared to your average GPU, which is great for clustering. The thing that's killing mainstream adoption for both options in the HPC sector is lack of double-precision (64-bit) floating-point support. There's a version of Cell on its way that fixes that problem, however. Note: I've not written code for the Cell myself, although I have for GPUs.