M³x: Autonomous Accelerators via Context-Enabled Fast-Path Communication
summary
M³x gives fixed-function accelerators autonomous access to OS services by adding a per-unit DTU, an Accelerator Support Module that implements a UNIX-like 5 protocol for memory and message endpoints, and a fallback slow-path for misrouted messages. Evaluation uses gem5 simulation and shows 4.7x speedup for PCIe-attached image pipelines and 30x lower CPU use versus CPU-assisted operation. Context switches serialize activity IDs via DTUs and keep fast-path messaging unless the kernel must schedule a receiver.
AI-written summary · about