# NVIDIA/cuda-tile

CUDA Tile IR is an MLIR-based intermediate representation and compiler infrastructure for CUDA kernel optimization, focusing on tile-based computation patterns and optimizations targeting NVIDIA tensor core units.

Repository: https://github.com/NVIDIA/cuda-tile
Canonical: https://ross.abutalabs.com/products/cuda-tile
Language: C++
License: NOASSERTION
License Family: other
Last push: 2026-08-24T08:04:29+00:00

## Health v2 (maintenance only)
Score: 82/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 94, longevity 21
- inputs: {"age_days": 301, "days_push": 9, "days_rel": 42, "gap_med": 6, "n_releases_24m": 14}
- flags: no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1015, forks 88 (observed 2026-08-28T04:03:14.079061+00:00)

## What it is
CUDA Tile IR is an MLIR-based intermediate representation and compiler infrastructure for optimizing CUDA kernels with tile-based computation patterns targeting NVIDIA tensor cores. It includes an MLIR dialect, Python bindings, bytecode serialization, and a conformance test suite.

## Use cases
- build and optimize CUDA kernels for tensor cores
- express tiled computations on NVIDIA GPUs
- programmatically construct and transform GPU IR in Python
- manage memory hierarchy in CUDA kernel development
- serialize tile IR to a binary bytecode format

## When to choose
- you develop high-performance CUDA kernels using tiling patterns
- you need an MLIR-based IR targeting NVIDIA tensor cores
- you want Python APIs for IR construction and transformation

## When to avoid
- you need a general-purpose CUDA compiler rather than a tile-focused IR
- you target non-NVIDIA GPUs
- you want a ready-to-use kernel library rather than compiler infrastructure

## Facets
- artifact type: library
- maturity: active
- function: compiler, gpu-computing, machine-learning
- domain: compilers, gpu-computing, developer-tools, machine-learning
- platform: cpp, python
- tags: mlir, cuda, tensor-cores, kernel-optimization, tiling, ir, nvidia, linux, gpu

## Member repositories
- NVIDIA/cuda-tile (main) score 82

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:14.079061+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:11:23.953146+00:00, confidence not recorded.
  - readme: https://github.com/NVIDIA/cuda-tile (fetched 2026-08-28T04:03:14.079061+00:00, sha 47a140fa0de7)
- Data as of 2026-08-30T08:39:29.467469+00:00.