alephnull on 13 June 2008
| Squilliam said: Hmm I was curious, does anyone know much about CUDA? I don't have time to work it out, but I think thats where the biggest competition to the Cell is going to come from. It seems Nvidia and IBM have similar goals but vastly different architectures. A modern RV770 chip does what, 1-1.2Teraflop FP single precision, probably 100gflop+ double vs a PS3 cell that does 250ish single precision... The question, can the Cell architecture in a cluster compete with future GPU designs as they become more general purpose? |
Despite what NVIDIA would have you believe CUDA does not cure cancer. The cell's SPE's work pretty much like normal cores only with 128k to work with and no coherency between other SPE's. You get the performance you are talking about only when you are in a situation in which it is suitable to run a single instruction on mutilple streams of data. This is perfect for edge detection for example. But terribly limited.
And quite frankly you will get significantly better performance out of an FPGA for the same types of tasks with more flexibility.
