Accelerated Software for a Post-Moore World
summary
Argues for replacing kernel‑first tuning with a system‑level optimizing compiler that treats a full workload as the unit of performance, integrates with PyTorch, and drives a runtime that does fine‑grain data movement and overlap of communication and computation. Presents a concrete implementation called Tachion that emits device‑specific object code from a parallel graph IR and DSL, reporting 1.4–2× kernel‑level speedups in early tests and a beta planned for summer 2026.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:44Modern software stack
- 2:09Post-Moore complexity drivers
- 3:16Kernels as assembly
- 5:04Kernels fail in memory-bound
- 6:53Cost of idle compute
- 7:16Compilers for parallel systems
- 9:45Tachion compiler
related talks
36mDL Compiler Panel Discussion
25mRace to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen...
19mCausalTuner: Feature-Aware Causal Guidance for Compiler Auto-tuning
56mAutomatically Scalable Computation
10mLightning Talk: Faster Than SOTA Kernels in Torch.compile With Subgrap...
25mHardware & Co-Design
10mMaximizing Kernel Development Productivity Under Performance Constraints
26mKeynote: Artificial Innovation Accelerates Artificial Intelligence