OpenTPU – An open-source AI accelerator, developed by AI

	OpenTPU – An open-source AI accelerator, developed by AI(github.com)
	337 points by fsbonetto 7 days ago \| 394 comments
	tl;dr: OpenTPU is an open-source AI accelerator designed largely by AI agents, running on a Xilinx Kintex-7 FPGA card with the entire stack (SystemVerilog RTL, ISA, simulator, compiler, and host software) in a single readable monorepo. It executes 10+ modern LLMs (Qwen3, LFM2, Gemma 4, Phi-4-mini, SmolLM3) with real weights, achieving bit-exact parity with its simulator and hitting 82-94% of DDR3 peak bandwidth during decode. Larger MoE models like Qwen3.5-35B run via PCIe expert streaming from the host.
	HN Discussion: ↑Excitement that this could mark a milestone toward AI singularity or recursive self-improvement ~Ironic concern that recursive self-improvement is now happening despite earlier warnings •Questions why frontier labs haven't already burned models into dedicated chips ↓Skepticism about implementation quality, noting incorrect floating point math in the RTL ↑Speculation about next steps, like AI designing its own hardware or FPGA-optimized architectures