35 d ago
openTPU: AI agents build an LLM accelerator that runs on a Kintex-7 FPGA card
Felipe Sens Bonetto used AI agents to write openTPU, an LLM inference accelerator with its own instruction set, simulator, compiler and host software. On an Inspur PCIe card with a Xilinx Kintex-7 and 2 DDR3 channels it runs 10 models, including Qwen3 and Gemma 4, with their real weights. A README GIF shows the chat demo running on the card.