IBM makes Spyre a first-class PyTorch backend

IBM said PyTorch ecosystem updates in the second half of 2026 turned its Spyre dataflow accelerator into a first-class PyTorch backend. The torch-spyre backend moved from IBM's in-house sendnn stack (~273K lines) to an Inductor extension (~38.6K lines), delivering 1.7x faster prefill and 2.4x faster decode on Granite 3.3 8B.

• Spyre is a 75W PCIe accelerator with 32 cores, 2 MB scratchpad each and up to 128 GB LPDDR5
• The torch-spyre backend cut code roughly 7x, from 273K to 38.6K lines
• On Granite 3.3 8B, prefill sped up 1.7x and decode 2.4x
• Softmax kernel time fell from 32.51 to 23.67 microseconds via scratchpad-aware scheduling