openTPU

GitHub - FeSens/openTPU: An open-source AI accelerator, developed by AI: RTL, ISA, simulator, compiler and profiler in one repo. Runs Qwen3, LFM2.5 and Qwen3.5 on a Kintex-7 PCIe card.

I’m using Claude Code to build an AI accelerator. It now runs Qwen3.5 on an FPGA. I wanted to learn more about AI accelerator design, so I started openTPU: an open-source accelerator developed with AI coding agents, including Claude Code. It already runs Qwen3.5 and other open-weight models on a Kintex-7 FPGA card. The part that's interesting is the hardware optimization workflow: - An agent reads the hardware code, timing reports, and lessons from previous attempts, then proposes a change. - Another agent implements it. - Automated checks compare the hardware against a Python simulator, test performance, and run simulations to measure the result. The biggest lesson so far has been giving the agents useful feedback: exactly which test failed, where the timing bottleneck is, and whether a change actually improved anything. Keeping that loop cheap makes experimentation much easier. I still have to nudge the larger architectural changes. Verification also has blind spots, we found floating-point bugs that our LLM inference tests hadn’t caught. Passing those tests only tells you so much. The whole stack is in one repo: hardware code, instruction set, simulator, a compiler for Python-style kernels, a profiler that shows what each hardware unit does cycle by cycle, and host tools for chat and monitoring. It’s free and open source, and you don’t need an FPGA to try it. You can run models through the simulator on a laptop and explore how the accelerator works. FeSens/openTPU Anyone else here using Claude Code for hardware design? I’d be interested to compare workflows and hear how you verify its changes.

Claude made a chip that runs Claude itself. Sparks of RSI FeSens/openTPU

openTPU An open-source AI accelerator, developed by AI: RTL, ISA, simulator, compiler and profiler in one repo. Runs Qwen3, LFM2.5 and Qwen3.5 on a Kintex-7 PCIe card. FeSens/openTPU

openTPU An open-source AI accelerator, developed by AI: RTL, ISA, simulator, compiler and profiler in one repo. Runs Qwen3, LFM2.5 and Qwen3.5 on a Kintex-7 PCIe card. FeSens/openTPU

🔥 Своя Canva на сервере, ускоритель ИИ на FPGA и ещё 1 - находки GitHub #shorts #github #opensource Больше полезных находок каждый день - в Telegram: t.me Подборка интересных находок с GitHub за 13.10.2026: shaongitbd/doory, FeSens/openTPU, storytold/wordcraft. Сохраняй, чтобы не потерять 💾 shaongitbd/doory: shaongitbd/doory Открытая альтернатива Canva на TypeScript: редактор для создания дизайнов и графики, который можно развернуть у себя. Пригодится тем, кому нужен свой визуальный редактор без привязки к облачному сервису, и разработчикам, которые хотят встроить подобный функционал в продукт. FeSens/openTPU: FeSens/openTPU Открытый ИИ-ускоритель, который почти целиком разработан с помощью ИИ: RTL на Verilog, набор инструкций, симулятор, компилятор и профайлер в одном репозитории. Запускает Qwen3, LFM2.5 и Qwen3.5 на PCIe-плате Kintex-7; будет интересен тем, кто изучает аппаратное ускорение нейросетей и FPGA. storytold/wordcraft: storytold/wordcraft Открытый текстовый редактор в духе Microsoft Word, написанный с нуля на Rust: привычная лента, стили, таблицы и рецензирование, читает и сохраняет .docx. Работает на macOS, Windows, Linux и в браузере; проекту неделя, так что пока скорее посмотреть, чем переезжать. #shorts #dev #github #useful #opensource гитхаб, гитхаб подборка, лучшее на гитхаб, новое гитхаб, гитхаб бесплатно, гитхаб ии, гитхаб мсп, гитхаб репы, гитхаб репозитории, лучшие репозитории, рейтинг гитхаб, github ai, best github, new github

Hop in, we are building OpenTPU! Hey everyone, I’m more of an FPGA guy than an ASIC guy. We’re building openTPU, an open-source AI accelerator with the whole stack in one repo: architecture, RTL, ISA, simulator, compiler, and host tools. It already runs some pretty respectable models like Qwen3.5 on a Kintex-7 PCIe card. With help from the community, we’ve now routed our first design using Nangate45. I’d love to get some ASIC eyes on it, especially on the floorplan, timing constraints, and anything we might be overlooking as we move from FPGA to ASIC. Feedback, contributions, and a good technical roast are all welcome. If you love this stuff, come have fun with us. FeSens/openTPU

HN 上 250 分，308 条评论，评论数比分数还多。 这条帖子讲的是 OpenTPU：一个开源 AI 加速器，标题写明是"由 AI 开发的"。AI 写软件大家已经见惯了，这回换成芯片，分量不一样。软件写错了，改一行重新跑就行；硬件错了，代价要大得多。 我是计算机工程 PhD 出身，看到这类项目，先想的是验证：AI 做出来的设计，别人能不能独立确认它确实能用。仓库是开源的，这一点谁都可以自己去查。 OpenTPU 的仓库在这里，设计文件都能直接看：FeSens/openTPU

AI agents designed hardware that runs Qwen on an old FPGA FeSens/openTPU. @aviflombaum sees a bigger story: what tiny, fast AI hardware could mean for medical devices. t.co

@AkitaOnRails Da uma olhada no projeto desse brazuka então: FeSens/openTPU Rodando LLMs em uma arquitetura própria em FOGAs