leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

v4.4 update. (#2979)

Browse Source
This commit is contained in:
Junkai-Wu
2026-01-25 00:46:17 +08:00
committed by GitHub
parent 2fafefb7b9
commit 9fba3195f9
293 changed files with 46344 additions and 2996 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
media/docs/cpp/blackwell_functionality.md
View File

@@ -131,7 +131,7 @@ instructions supported by CUTLASS.
|[8](#bs_rows_4_5_7_8_10) | Dense | mx_float6_t | mx_float6_t | TN, NN, NT, TT | 128 | 128 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf6_void_bf16_tn_layout.cu)<br>[NT unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf6_void_bf16_nt_layout.cu)|
|[9](#bs_rows_6_9_11) | Dense | mx_float6_t | mx_float8_t | TN, NN, NT, TT | 128 | 16 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf8_void_f32_tn_layout.cu)<br>[NT unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf6_mxf8_void_f32_nt_layout.cu)|
|[10](#bs_rows_4_5_7_8_10) | Dense | mx_float8_t | mx_float6_t | TN, NN, NT, TT | 16 | 128 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf6_f16_f8_tn_layout.cu)<br>[NT unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf6_f16_f8_nt_layout.cu)|
|[11](#bs_rows_6_9_11) | Dense | mx_float8_t | mx_float8_t | TN, NN, NT, TT | 16 | 16 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_tn_layout.cu.cu)<br>[NT unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_nt_layout.cu)|
|[11](#bs_rows_6_9_11) | Dense | mx_float8_t | mx_float8_t | TN, NN, NT, TT | 16 | 16 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_tn_layout.cu)<br>[NT unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_tensorop_gemm/mxf8_mxf8_void_f8_nt_layout.cu)|
|[12](#bs_rows_1) | Sparse | nv_float4_t | nv_float4_t | TN | 32 (N) / 64 (T) | 32 | mxf4nvf4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_sparse_tensorop_gemm/sm100_bssp_gemm_nvf4_nvf4_f32_void_f16_o_tnn.cu) |
|[13](#bs_rows_2) | Sparse | mx_float4_t | mx_float4_t | TN | 32 (N) / 64 (T) | 32 | mxf4, mxf4nvf4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_sparse_tensorop_gemm/sm100_bssp_gemm_mxf4_mxf4_f32_f16_f16_o_tnn.cu) |
|[14](#bs_rows_3) | Sparse | mx_float4_t | mx_float4_t | TN, NN, NT, TT | 128 (N) / 256 (T) | 128 | mxf8f6f4 |[TN unit tests](https://github.com/NVIDIA/cutlass/tree/main/test/unit/gemm/device/sm100_blockscaled_sparse_tensorop_gemm/sm100_bssp_gemm_mxf4_mxf4_f32_f16_f16_q_tnt.cu) |
@@ -652,7 +652,7 @@ auto tensor_sfb = make_tensor(bptr, layout_sfb);
auto val_a_mk = tensor_sfa(make_coord(m,k,0));
```
# Blackwell SM120 GEMMs
The NVIDIA RTX 5000 Series GPUs introduce support for new narrow precision (4bit and 6bit) block-scaled and non-block-scaled tensor cores. The PTX ISA has extended the `mma` instructions to support these data formats which are 1x to 4x faster than Ada architecture's fp8 tensor cores. For more detailed information see [`mma` PTX documentation](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#multiply-and-accumulate-instruction-mma).
The NVIDIA RTX 5000 Series GPUs introduce support for new narrow precision (4bit and 6bit) block-scaled and non-block-scaled tensor cores. The PTX ISA has extended the `mma` instructions to support these data formats which are 1x to 4x faster than Ada architecture's fp8 tensor cores. For more detailed information see [`mma` PTX documentation](https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-for-mma).
CUTLASS 4.0 has added support for these newly introduced narrow precision GEMMs. Similar to the Blackwell SM100 GEMMs, the SM120 GEMMs can be built using the collective builder interface. See examples in [examples/79_blackwell_geforce_gemm/](../../examples/79_blackwell_geforce_gemm/) and unit tests listed below.