Functionality#

For dependency version requirements, refer to the Quick Start Guide section.

Supported MMA Operations#

NVIDIA Ampere Architecture:

  • FP16 / BF16 tensor core instructions

NVIDIA Hopper Architecture:

  • FP16 / BF16

  • FP8

NVIDIA Blackwell Architecture:

  • FP16 / BF16

  • TF32

  • I8

  • F8

Notable Limitations#

For current constraints and unsupported features, refer to the Limitations section.