Leveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
summary
A compilation framework converts Triton kernels into standalone shared objects with C APIs, removing Python/PyTorch dependencies and loading PTX/cubin/fatbin at runtime. Replaced kernels in Lulesh and a tensor contraction library run on NVIDIA and AMD GPUs with near‑parity performance (often within ±10%) for tile‑friendly kernels, while XSBench’s irregular workload shows a large slowdown. Kernel fusion and ahead‑of‑time PTX emission prove valuable; C ABI shared objects and kernel fusion across Triton kernels enable practical HPC integration.
AI-written summary · about
chapters (12)
- 0:00Portable GPU kernels challenge
- 1:27Limitations of portability methods
- 3:34Triton for HPC kernels
- 8:23Triton in HPC environments
- 11:17Removing Python for HPC
- 13:22Debugging Triton trade-offs
- 15:04Triton kernels in Lulesh
- 16:48Triton libraries for HPC
- 17:52Triton in quantum chemistry
- 19:39Triton for XSBench
- 21:31Triton HPC future limits
- 22:56Q&A
related talks
08mGPU Programming with Triton Kernels
10mMaximizing Kernel Development Productivity Under Performance Constraints
05mLightning Talk: GPU Programming with C++ and Triton
16mLightning Talk: Triton Compiler
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
24mHelion 1.0: A High-Level DSL for Performance Portable Kernels
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
24mPortable High‑Performance LLM Serving: A Triton Backend for...