Race to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
summary
Meta scales PyTorch into a hardware OS using Triton and Torch Inductor to target a family of MTIA accelerators with an MLIR Triton backend, production deployments on MTIA, and two optimisation paths: extended Triton language and a C++ low level model for roofline performance. Claims include Triton for MTIA enabling GPU Triton kernels to port with minimal changes and 120 PyTorch A10 ops in under 10 hours from AI kernel generation. Goal: day‑zero support for new silicon with fungibility, velocity, and performance.
AI-written summary · about
chapters (12)
- 0:00Low-friction AI infrastructure
- 1:28AI workload and hardware diversity
- 3:15Meta's MTIA accelerator
- 6:05Hardware diversity in software
- 9:19PyTorch for frameworks and hardware
- 10:24TorchInductor and Triton
- 13:54Expanding operator coverage
- 16:01Peak performance with low-level control
- 18:37Triton in production at Meta
- 19:44AI-assisted kernel generation
- 22:14Future AI infrastructure
- 23:17Heterogeneous hardware closing
related talks
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR
12mLightning Talk: PyTorch 2.0 on the ROCm Platform
25mMoving Inference to Triton Servers
16mLightning Talk: Triton Compiler
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
25mHardware & Co-Design
28mHelion: A High-level DSL for Kernel Authoring
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler