openTPU: A One-Person Open-Source AI Accelerator Designed by AI Agents, Running 10 LLMs on an FPGA
ai · x · 2026-10-07
Developer BonettoSens released openTPU, an open-source AI accelerator largely designed by AI agents — described as a one-person project doing team-scale work.
- Full stack in one repo: RTL, custom ISA + compiler, Python simulator (Verilator), profiler, and FPGA implementation (Apache-2.0)
- Runs 10 modern LLMs (Qwen3, LFM2.5, etc.) on a Kintex-7 PCIe card, producing real tokens, with otpu-chat and otpu-smi showing utilization and DRAM bandwidth
- Current design targets Kintex-7 DSPs/BRAM and DDR3 + LiteDRAM, closing timing around 133 MHz (WNS +0.032 ns)
- Kevin Gubbi notes RTL writing is getting easier but silicon is the hard part; moving to ASIC/SoC is a big leap, though the ISA/compiler ideas carry over
- The project asks how far agents can go at hardware design — and whether they can build the chip that runs their own inference
More from coding & agent
- Pre-registered study: selection beats extraction in agent memory, 3061x cheaper to write — Rishabh Sharma · 2026-10-07
- Cursor iOS app adds remote control for local coding agents, on by default — tetsuoai · 2026-10-07
- Claude Code v2.1.292 adds effort-level control for subagents — lydiahallie · 2026-10-07
- 100B decisions in 5 minutes: a layered agent pipeline where Opus only probes 96 hotspots — Arindam_1729 · 2026-10-07
- requarks-wiki-mcp: 29-Tool MCP Server Lets AI Agents Query Wiki.js Knowledge Bases — modelcontextprotocol · 2026-10-07
- Cursor's Origin API ships four breaking changes, email fields now scope-gated — tetsuoai · 2026-10-07