IBM moves Spyre accelerator to PyTorch Inductor, cutting compiler code 7x
PyTorch · x · 2026-10-13
IBM engineers detail on the PyTorch blog how they migrated the torch-spyre compile backend from their from-scratch sendnn stack to an open-source backend built on PyTorch Inductor.
- Spyre is IBM's dataflow AI accelerator for enterprise inference: 75W, PCIe-attached, 32 small cores on a ring with 2 MB local scratchpads, no shared cache, data moved in 128-byte "sticks"
- Most of Inductor carried over untouched; effort focused on data movement, memory management, and work distribution where Spyre diverges from GPU assumptions, yielding a compiler 7x smaller
- Key difference: on GPUs hardware caches keep hot values close; on Spyre the compiler must decide what stays close, concentrating work in lower-level codegen
- The post demonstrates Inductor's extensibility for bringing new non-GPU hardware into PyTorch
More from Infra
- Basalt: A fast Blackwell inference engine for Qwen 3.8 Flash Next with dynamic vision — jesdga95 · 2026-10-13
- 80B MoE Model Runs in a Chrome Tab: llama.cpp WebGPU Backed Pages Experts From Disk — naklitechie · 2026-10-13
- The End of the Best-Model-Wins Era: AI Value Is Rotating Into Data Centers and Apps — RihardJarc · 2026-10-13
- Market rotating out of semis into hyperscalers and application layer, analyst says trend will continue — RihardJarc · 2026-10-13
- Intel hikes CPU prices 15-20% across Panther Lake and older lines; AMD expected to follow — TheZachMueller · 2026-10-13
- Mistral models head to orbit in 50-satellite AI constellation with NVIDIA compute — MistralAI · 2026-10-13