3 ms·
Not quite. This is FP16, supercomputers are typically benchmarked using double precision flops, so FP64. So roll in a factor of ~4 there, or a couple of years.
by martinpw 8y ago
Not quite. This is FP16, supercomputers are typically benchmarked using double precision flops, so FP64. So roll in a factor of ~4 there, or a couple of years.
- johndough 8y agoFP16 has 11 bit mantiassa, whereas FP64 has 53 bits. Integer multiplication scales quadratically with the number of bits, so a factor of ~23 might be more accurate.