AI Designs an Open Inference Accelerator
OpenTPU presents an unusually complete open accelerator stack whose RTL, instruction set, simulator, compiler and profiler were developed with AI agents, and reports running several small open models on an FPGA card. The important weak signal is not another accelerator project but AI crossing from software generation into a tightly coupled hardware-software design loop. One repository cannot establish reproducibility, performance or a market; independent synthesis results, boards and downstream contributors are required before publication.