Tensor Cores Optimization

Warp Matrix Multiply-Accumulate (WMMA)

CUDA API allowing warps of 32 threads to efficiently perform matrix multiply-accumulate operations directly on Tensor Cores with access to fragmented registers.

← Back