Redfire75369’s avatarRedfire75369’s Twitter Archive—№ 3,519

    1. With the Biren and new Intel disclosures on their HPC/AI Data Center GPUs, the table of peak compute throughputs has been updated. I also added Xeon Phi, PEZY, NEC and Preferred Networks for no reason.
  1. …in reply to @Redfire75369
    Ponte Vecchio's new specs barely edge out H100 PCIe (though at higher power). BR100 claims to beat basically everything, but it has about 1/5 the memory bandwidth to compute ratio of the MI250X, meaning its highly bandwidth-bound.