NVIDIA/cutlass

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

10317 stars · 2046 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

CUTLASS is NVIDIA's collection of CUDA C++ template abstractions and Python DSLs for implementing high-performance GEMM and related linear algebra computations on NVIDIA GPUs. It provides reusable, tunable kernel building blocks supporting many data types across Volta through Blackwell architectures.

Use cases

• write custom high-performance GEMM kernels for NVIDIA GPUs
• prototype CUDA kernels in Python without deep C++ expertise
• implement mixed-precision matrix multiplication with FP16, BF16, FP8, or FP4 types
• integrate optimized tensor core kernels into deep learning frameworks
• tune tiling sizes and data movement strategies for GPU performance

When to choose

• you need maximum GEMM or linear algebra performance on NVIDIA tensor cores
• you are developing custom CUDA kernels for deep learning or HPC
• you want a Python DSL for GPU kernel programming with C++-level performance
• you need support for the latest NVIDIA architectures and low-precision data types

When to avoid

• you just need a drop-in BLAS library rather than kernel building blocks (use cuBLAS)
• you target non-NVIDIA GPUs or CPUs
• you want a simple high-level API without performance tuning

Facets

library · maturity active

machine-learning deep-learning gpu-computing math compiler deep-learning gpu-computing machine-learning performance cpp python windows cross-platform cuda gemm linear-algebra tensor-cores kernels cute-dsl nvidia template-library algorithms gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVIDIA/cutlass")