sharing my first blog! a meta harness for self-improving AI chip design…
check it out if u wanna see how RL could work in chip design task, or how an ASIC design can become a hillclimbing task
it introduced ASIC vs. GPU, RL Env setup, and the best generated design can run Kimi K3 at 87000 tps theoretically, which is 10x faster than the example shared by Kimi Team
selected design files are open-sourced, happy to share the harness code with the enthusiasts! extremely enticing to bring RL capabilities to more complex system
luoluo.ai/blog/kimi-k3