Unveiling MakoGenerate: Lightning-Fast AI-Driven GPU Kernel Creation in Just 60 Seconds!

MakoGenerate, now available in research preview, is an innovative AI agent designed to autonomously generate, compile, validate, and benchmark GPU code, specifically for CUDA and Triton on Nvidia GPUs from Hopper to Blackwell. Users can quickly start by creating an account and selecting a problem from a dropdown menu, generating their first GPU kernel in under a minute. The custom prompt feature allows users to provide specific contexts and optimization techniques. MakoGenerate combines large language model (LLM) code generation with an automated feedback loop and an optional evolutionary search for high-performance kernel development. This approach refines kernels through iterative improvements, producing multiple candidates for comparison, which are validated and benchmarked. The program also features an early access initiative for those interested in enhanced performance capabilities. Users can join a Discord community to share results and collaborate. For early access, visit mako.dev and complete the Contact Us form.

Source link