Hacker News new | ask | show | jobs
by pezezin 813 days ago
"Exceeds" is an understatement. The Cell CPU had a maximum performance of 200 GFLOPS of FP32. A Nvidia 4090 has 73 TFLOPS of FP32, equivalent to 360 PS3.
1 comments

For physics simulations they are using double precision. They increased the cluster to 400 PS3s so it continued to be useful in the 2010's.

https://web.uri.edu/gravity/ps3/

But I guess the cluster overhead would reduce performance to less than a 4090 in real applications.

Oh, in that case consumer GPUs absolutely suck at FP64, the vendors want you to buy the expensive data-center version. The RTX 4090 has a 1/64 rate when computing with FP64, so only 1.2 TFLOPS!

But from what I can find, the Cell also sucked at FP64, with a rate of 1/10 for a total of 15 GFLOPS (https://en.wikipedia.org/wiki/PlayStation_3_technical_specif... , second paragraph). The 400 PS3 cluster would be 6 TFLOPS or 5x RTX 4090.