Maximizing Kernel Development Productivity Under Performance Constraints
summary
Triton is presented as a mid‑level kernel language that trades some CUDA flexibility for much faster iteration and easier maintenance, supporting pointers in tensors and an SRAM‑style tensor model but requiring tile shapes that are powers of two along each dimension. With kernel tweaks it achieves faster than cuBLAS on H100 with FP8 for large matrices and matches handwritten kernels on MI300/Intel in many cases. It shifts maintenance cost from kernel engineers to compiler work and lets researchers write production kernels directly.
AI-written summary · about
chapters (10)
- 0:00Triton introduction
- 1:06CUDA and graph compiler limits
- 2:30Triton as middle ground
- 3:58Triton performance trade-offs
- 4:39Triton SRAM and pointers
- 5:41Matrix multiplication in Triton
- 6:25Triton kernel improvements
- 7:07Triton at OpenAI
- 8:53Reducing tech debt with Triton
- 9:34Triton vs CUDA/assembly
related talks
08mGPU Programming with Triton Kernels
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
16mLightning Talk: Triton Compiler
05mLightning Talk: GPU Programming with C++ and Triton
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
24mPortable High‑Performance LLM Serving: A Triton Backend for...
28mHelion: A High-level DSL for Kernel Authoring
20mUnlocking Performance: Harnessing LLMs To Streamline GPU Kernel Development in...