Comment by saagarjha
There's a 2x performance hit from the weird restriction on fp32 accumulation, plus the fact that 5090 has "fake" Blackwell (no tcgen05) which limits the size and throughput of matrix multiplication through the tensor cores.
There's a 2x performance hit from the weird restriction on fp32 accumulation, plus the fact that 5090 has "fake" Blackwell (no tcgen05) which limits the size and throughput of matrix multiplication through the tensor cores.