An AI-designed open-source TPU runs Qwen3 and other LLMs on a $200 FPGA card

AI is now capable of developing its own inference hardware

openTPU is an open-source AI accelerator whose RTL, ISA, simulator, compiler, and profiler were developed by AI agents. It runs ten modern models, including Qwen3 and LFM2.5, on a Kintex-7 PCIe card, matching the simulator bit for bit. Decode speeds reach 85.8 tokens/s for LFM2.5-230M, with DRAM bandwidth at up to 94% of peak. The project explores how far AI can go in hardware design and whether it can build the chip that runs its own inference.