KernelBot: A Competition Platform for Writing Heterogeneous GPU Code

Alex L Zhang, Matej Sirovatka, Erik Schultheis, Benjamin Horowitz, Mark Saroufim

Keywords: gpu, oss, open-source, competition, accelerator, mlsys

TL;DR: A fully open-source competition for sourcing high-quality GPU code spanning many different hardware vendors.

Abstract: Writing performant code for GPUs and parallel processors is critical to building many machine learning systems. However, writing these kernels requires significant manual effort and expertise. Furthermore, there is a lack of public, high-quality kernels to learn from. To address these issues, we introduce KernelBot, a community-based open-source online platform for hosting code competitions where users optimize GPU algorithms. At the end of every competition, we open source all participant submissions under a permissive license. In this paper, we discuss how we designed the platform with the goal of making the submission process interactive, minimizing barriers to entry such as cold starts. Additionally, we built the competition such that top-performing kernels could be taken as-is to accelerate popular models. Since launching in early March 2025, our platform has received over 25K submissions in the span of two competitions. The top kernels in these competitions have also had real impact and success on two popular hardware vendors, NVIDIA and AMD.

Submission Number: 18