github.comgithub.comgithub.com Visit site

This Open Source AGENTIC RISC-V processor completed an AI processing heuristic on GPT-2 with over ~50x improvement in efficiency. The core contains all the features of a modern processor with integrated AI (similar to integrated graphics) which is scalable in cores, processing lanes and cache. This means a potential processor dev can have an 8-core CPU with 64MB L3 cache and GeForce-like AI throughput by changing a few configuration variables, generating the OpenRoad mask for photolithography and have it produced within 12 weeks, given it still boots linux (openwrt) in the given QEMU. The project also supplies a fully open source RISC-V bios that runs on a web UI (svelteD), and everything needed to optimize the speed. Integrated graphics are under heavy development.

Similar optimizations (resident weights) on the AI workloads have been hypothesized and explored to an extent in https://github.com/ggml-org/llama.cpp/discussions/24528 and https://github.com/ggml-org/llama.cpp/discussions/28248

This implementation in LibreCore includes many other conceptual improvements regarding neural networks and it is the first time it has been achieved in hardware.

Because this is an open source RISC-V processor, anyone can boot up Grok Build or Claude (perhaps in Devin IDE) and start developing state-of-the-art nanotechnologies mostly for Free!