Datatype and hardware support matrix#
This page summarizes which low-precision quantization formats Transformer Engine supports on NVIDIA GPU architectures (by compute capability).
Compute capability reference#
Compute capability |
Architecture |
Representative GPUs |
|---|---|---|
8.0, 8.6 |
Ampere |
A100, A10, A40, RTX 30 series |
8.9 |
Ada Lovelace |
L4, L40S, RTX 40 series |
9.0 |
Hopper |
H100, H200 |
10.0, 10.3 |
Blackwell (data center) |
B200, B300, GB300 |
12.0 |
Blackwell (workstation / consumer) |
RTX PRO 6000, RTX 50 series |
The architecture and GPU columns show representative examples. For a complete list, see NVIDIA’s CUDA GPU compute capability list.
Format support by compute capability#
Compute capability |
BF16 |
FP8 (per tensor) |
FP8 block scaling |
MXFP8 |
NVFP4 |
|---|---|---|---|---|---|
8.0, 8.6 (Ampere) |
Yes |
No |
No |
No |
No |
8.9 (Ada) |
Yes |
Yes [1] |
No |
No |
No |
9.0 (Hopper) |
Yes |
Yes |
Yes [2] |
No |
No |
10.0, 10.3 (Blackwell DC) |
Yes |
Yes |
Yes [2] |
Yes |
Yes |
12.0 (Blackwell workstation) |
Yes |
Yes |
Yes [2] |
No [3] |
Partial [4] |
BF16 requires compute capability 8.0 or higher.
FP8 (per tensor) covers the
DelayedScalingandFloat8CurrentScalingrecipes. It requires compute capability 8.9 or higher.FP8 block scaling is the
Float8BlockScalingrecipe.MXFP8 is the
MXFP8BlockScalingrecipe.NVFP4 is the
NVFP4BlockScalingrecipe.
Default recipe by architecture#
When no recipe is passed explicitly, Transformer Engine selects a default based on the device:
Compute capability |
Default recipe |
|---|---|
8.9, 9.0 (Ada, Hopper) |
|
10.0, 10.3 (Blackwell DC) |
|
12.0 (Blackwell workstation) |
|