4 ms·
A new, undocumented INT1 tensor core instruction in Nvidia Ampere
- lmh 6y agoFor the hardware and architecture nerds, CUTLASS now appears to include support for an INT1 (binary) AND-popcount matrix multiplication, in addition to the XOR-popcount present in Turing.