Can you do bf16 x bf16 -> FP32 matmul?

1

Possibly a noob question here. Say I have two bf16 tensors. Is there a way to do matmul on them and return the accumulator in FP32?

2

This might be implementation dependent, but in almost all cases accumulation will be done in fp32 by default. You can try out variant that might work

1 Like