openTPU, an open-source accelerator built with AI agents, runs language models on an FPGA card
Released by developer Felipe Sens Bonetto under the Apache 2.0 license, openTPU bills itself as an open-source AI accelerator “developed by AI.” One repo holds the hardware design, instruction set, simulator, compiler and card software. It runs models like Qwen3, Gemma 4 and Phi-4-mini on an Inspur YPCB-00338 PCIe card with a Kintex-7 chip. Per the README, with 4-bit weights Qwen3-0.6B generates 30.7 tokens per second and LFM2.5-230M 82.1. Qwen3.5-35B-A3B, too big for the card, reaches 3.95 by streaming experts from the host.
The project drew 348 points and 405 comments on Hacker News on October 6. Bonetto says the same AI loop was used earlier to develop RISC-V CPU cores; the README doesn’t say how the work split between agents and humans. On Habr, a user ran it on a card with the same chip bought for about $50. The design runs only on FPGAs so far, decoding speed is bound by DDR3 memory bandwidth, and it isn’t meant as a copy of Google’s TPU.
Sources
- GitHub (FeSens/openTPU), “openTPU”, README: card, models, speeds, limits, Apache 2.0 (github.com)
- Hacker News (fsbonetto), “OpenTPU – An open-source AI accelerator, developed by AI”, Launch thread, Bonetto's account of the method, criticism in comments, (news.ycombinator.com)
- GitHub (FeSens/auto-arch-tournament), Earlier project: agent loop that evolves a RISC-V core (github.com)
- GIGAZINE, “OpenTPU, an open-source AI accelerator developed using AI technology, has emerged.”, Independent coverage; not a copy of Google's TPU, (gigazine.net)
- Habr (Konstantin Tyutin), “Нафиг NVIDIA! Я запустил Qwen на FPGA за $50 — а свой TPU назвал бы «КЕДР». Часть 1”, Independent setup: card for about $50 from AliExpress, (habr.com)
About this story
Spotted an error in this story? [email protected] · Instagram
Short link: thejarrus.com/en/opentpu
This story in Turkish: Yapay zeka ajanlarıyla geliştirilen açık kaynak hızlandırıcı openTPU, dil modellerini bir FPGA kartında çalıştırıyor
On the same topic
Nvidia’s share of China’s data center AI chip market fell to 34% in the first half of 2026 as Huawei reached 28%, IDC data shows
An IDC report seen by Reuters shows Nvidia shipped 678,000 AI accelerators in China in the first half, down 44%, while Huawei shipped 572,000 units.
Linus Torvalds built an experimental Linux kernel patch series with Claude Code
Torvalds says Claude Code wrote the code, tests and first draft. One test ran 23 times faster, but the series isn’t merged and was handed to Kiryl Shutsemau.
Linus Torvalds published an experimental Linux kernel patch series that he built with Claude Code
Torvalds says the code, tests and first draft of the text of his 8-patch series to speed up page cache reads were done with Claude Code under his direction.