OpenTPU: New Open-Source AI Accelerator Developed for Enhanced Performance

Oct 6, 2026

AI Summary

OpenTPU is an open-source AI accelerator that integrates hardware design, a simulator, and software tools in a single repository. It successfully runs multiple AI models on a Kintex-7 PCIe card, demonstrating advancements in AI hardware design and efficiency.

OpenTPU combines RTL, ISA, a simulator, compiler, and profiler in one repository.
It is capable of running models such as Qwen3, LFM2.5, and Qwen3.5 on a Kintex-7 PCIe card.
The design allows for a comprehensive understanding of AI accelerator functioning, from hardware design to software execution.
OpenTPU has been tested with ten modern models, producing identical outputs on both the simulator and the physical card.
The accelerator utilizes a simple architecture where each data movement is an instruction, allowing for precise performance tracing.
It supports mixture-of-experts models, optimizing memory usage by streaming experts from host storage.
The project is designed for learning and encourages contributions, requiring only Python and Verilator for most development tasks.
OpenTPU is licensed under the Apache License 2.0.

open-sourceai acceleratorhardwaretechnologyinnovation

OpenTPU: New Open-Source AI Accelerator Developed for Enhanced Performance

Related Stories

Sophos reduces cyber-threat investigation time significantly using OpenAI Daybreak

Website Simulates Elizabeth Holmes' Desk with Court Documents from Theranos Trial

Jev, an AI for making quick decisions, has been a viral hit in Silicon Valley. But OpenAI is hot on its heels

Google introduces Gemini AI with task management capabilities for businesses