This past weekend, developers from across the AI and systems programming communities came together for Modular Hack Weekend: a global, virtual hackathon focused on GPU programming and model implementation with Mojo and MAX. Held primarily online through our Discord server and forum, this event brought fresh energy, bold ideas, and a powerful reminder of what this community can build in just 48 hours.
The hackathon kicked off with a hybrid GPU Programming Workshop on Friday evening, where attendees tuned in via livestream around the world, and many attended in-person at our office in Los Altos, California. The workshop featured lightning talks from:
Our workshop also explored the foundations of writing performant kernels in Mojo and how MAX makes it possible to build model graphs that run across heterogeneous hardware, and concluded with plenty of time for mingling. Finally, the hacking began, with participants meeting teammates both in-person at the workshop and online via the Discord server and forum.
This event wouldn’t have been possible without the support of our partners:
We’re incredibly thankful to all three partners for their support in making this hackathon a success.
GitHub repo and forum post
Martin is a mechatronics engineer who pivoted into software development. He works with Python and SQL professionally and is a major contributor to the Mojo standard library in his spare time. For Modular Hack Weekend, he chose to implement the Fast Fourier Transform on GPU using Mojo. The project built on his university background in signal processing and gave him the opportunity to explore parallel computing.
“This is my first time programming anything on a GPU beyond just using TensorFlow. As many people have said, it was surprisingly easy. I also decided to do something I'm familiar with and is solved in 1 dimension to avoid the many headaches that arise with multidimensional tensors. If so many people can manage to get up to speed in GPU programming in a weekend, it goes to show just what an incredible set of tools Modular is building!”
Martin walked away with first place and an RTX 5090.
Seth is a bioinformatics software engineer at Bio-Rad Laboratories with a strong interest in developer tooling and high-performance computing. For his project, he implemented Mojo-Lapper, a GPU-accelerated library for interval overlap detection. The kernel uses the BITS algorithm and achieved 60 to 140x speedups over CPU versions, with practical applications in genomics, databases, and time-series analysis.
“Mojo-Lapper is a library for doing interval set operations with a unified API for both CPU and GPU. As part of that I ported the BITS algorithm for counting interval intersections as a Mojo kernel for 140x speedup over a single-threaded CPU version. There’s still lots of performance to be gained here, but once again, it was impressive how code written for the CPU could pretty much just be dropped on a GPU thread for huge gains.”
Seth earned second place and an RTX 5080.
GitHub repo and forum post
Thomas is a computer science master’s student focused on AI and quantum computing. He previously built low-level systems like a C threading library and visualization tools using WebAssembly. For the hackathon, he created QLabs, a GPU-accelerated simulator for quantum circuits written entirely in Mojo. It was his first time doing any kind of GPU programming.
“Learning by doing was incredibly rewarding. The Mojo documentation and the GPU programming puzzles gave me a good foundation and helped me start strong. But more than that, the chance to engage with passionate, smart, and talented people, both from Modular and among participants, made the experience truly memorable.”
Thomas took third place and won an RTX 5070.
While only three prizes were awarded, several other projects stood out for their technical quality and creativity. Check out all the projects in our forum.
This project reimagines the Boids simulation using GPU programming in Mojo. By replacing a brute-force kernel with a spatially-aware pipeline, Quinn achieved a 14.5x speedup. The optimized version reduces complexity from O(N²) to O(N), enabling real-time simulation of 100,000 boids.
Whisper‑Mel‑Mojo is a portable Mojo kernel that fuses log-Mel spectrogram extraction with a 3×3 average convolution, replicating Whisper’s frontend. It runs entirely on-device with zero host-to-device transfers, keeping audio in GPU memory and reducing overhead.
We’re grateful to everyone who joined, shared ideas, and gave feedback during Modular Hack Weekend! You made it a fast-moving, high-energy event. Thanks again to our partners at NVIDIA, Lambda, and GPU MODE for supporting this effort and helping more people access modern compute infrastructure.
If you missed the GPU Programming Workshop, the recording is available on Modular’s YouTube channel. And if you're still hacking on your project or trying MAX and Mojo for the first time, the Modular forum and Discord are always open – we can't wait to see what you build!
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。