AI-assisted openTPU reportedly runs language models on a used FPGA card

Brian Roemmele reports that openTPU runs Qwen, Gemma and Phi models on a used FPGA card. The project credits AI agents with designing and tuning the accelerator, but its documentation leaves merges to a human; the exact split of work remains unclear. Its README says the simulator runs on a laptop, while using the physical card requires FPGA tools and a paid or evaluation Vivado license.