openTPU: an open-source inference accelerator for AI development
What happened
On October 7, Hacker News' front page covered openTPU, an open-source AI accelerator developed by AI. The project has run ten models on an FPGA card using real weights, with output tokens matching the simulator bit for bit, showing it can run model inference. The same repository holds the SystemVerilog hardware design, instruction set, simulator, kernel language, compiler and host software, and it discloses how device speed and end-to-end speed were measured.
Written by AI from the coverage · updated 3 hours ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageAI is now capable of developing its own inference hardware
openTPU 是由 AI 开发的开源 AI 加速器，已在 FPGA 卡上用真实权重运行十个模型，输出 token 与模拟器逐位一致。项目在同一仓库提供 SystemVerilog 硬件设计、指令集、模拟器、内核语言及编译器和主机软件，并披露设备与端到端速度的测量方法。
Heat over time
Not enough continuous observations to draw a trend yet.