Triton paper: An intermediate language and compiler for tiled neural network computation
paper
Fields
id
sources/triton-paper-an-intermediate-language-and-compiler-for-tiled-neuPublisher
ACM MAPL 2019
Kind
paper
Published
2019-06-22
Accessed
2026-10-11
notes
Original Triton paper by Tillet, Kung, Cox (Harvard). Describes Triton-C (C-like language), Triton-IR (LLVM-based IR), and Triton-JIT compiler. Demonstrates matmul and convolution on par with cuBLAS/cuDNN. The tile-centric approach is the key abstraction.
2026-10-11: Re-read full MAPL 2019 publication PDF (Tillet, Kung, Cox; DOI 10.1145/3315508.3329973). Claims hold: describes Triton-C, Triton-IR, and Triton-JIT with tiled compute abstractions achieving cuBLAS/cuDNN-level matmul/convolution performance on V100/Titan V. Raised 0.5 -> 0.85, status draft -> verified as canonical peer-reviewed publication.
Notes
Original Triton paper by Tillet, Kung, Cox (Harvard). Describes Triton-C (C-like language), Triton-IR (LLVM-based IR), and Triton-JIT compiler. Demonstrates matmul and convolution on par with cuBLAS/cuDNN. The tile-centric approach is the key abstraction.
2026-10-11: Re-read full MAPL 2019 publication PDF (Tillet, Kung, Cox; DOI 10.1145/3315508.3329973). Claims hold: describes Triton-C, Triton-IR, and Triton-JIT with tiled compute abstractions achieving cuBLAS/cuDNN-level matmul/convolution performance on V100/Titan V. Raised 0.5 -> 0.85, status draft -> verified as canonical peer-reviewed publication.
Cross-references
This record references no other records.