Modular GPU Kernel Hackathon

​Please note that you'll need to apply via AGI House here to attend this event.
​Development of AI/GPU Compute Kernels has long been constrained by legacy technologies–C++ with its memory management complexities, Python with its performance limitations, and vendor-specific platform like CUDA that are vendor specific. It's time for a breakthrough: Mojo 🔥 delivers that path forward, offering a modern language designed specifically for high-performance computing without the historical baggage.

At the Modular GPU Kernel Hackathon, you'll get hands-on experience using Mojo for kernel development. Discover how to write cleaner, faster, and more portable code for the latest NVIDIA and AMD GPUs that can transform your AI systems programming. Join fellow developers at the cutting edge of kernel innovation and be part of the movement redefining what's possible in high-performance computing!

Modular is thrilled to present this event at AGI House, in collaboration with event hardware sponsor Crusoe and our friends at GPU Mode.

Event Schedule 📅
10:00 AM – Doors open, early-bird networking (meet future teammates!)
11:00 AM – Keynote from Chris Lattner and talk from Mark Saroufim (GPU Mode and PyTorch)
12:00 PM – Hackers pitch projects
12:15 PM – Hacking starts and lunch
4:00 PM – Project check-in
6:00 PM – Dinner
8:00 PM – Demo presentations

Prizes 🏆
1st Prize: NVIDIA RTX 5090 GPU
2nd Prize: NVIDIA RTX 5080 GPU
3rd Prize: OLED Steam Deck

Project Inspiration 🤩
Traditionally, GPU programming has relied on proprietary, vendor-locked languages and unwieldy libraries that lack composability. Mojo offers a fresh perspective, providing developers with tools and libraries to create high-performance kernels in an intuitive way. This hackathon challenges you to use Mojo to program and optimize GPU kernels, with an opportunity to win prizes for your work.

During this hackathon, push the limits by implementing novel ML algorithms and kernels using Mojo.

Here are just a few examples of kernels you could write:
• Write a transformer attention block.
• Port a GPTQ implementation to NVIDIA H100 or AMD Mi300X
•Implement a Fast Fourier Transform (FFT) for GPU.
• Write a computationally fast cumulative sum (cumsum) operation.
• Write computationally fast image processing kernels such convolution or non-maximum suppression (NMS).
• Implement Radix sort on GPUs.

What will you build? 🚀

Rules
1. All attendees must contribute code to a public GitHub repository.
2. Teams must consist of four members or fewer and can be formed during the hackathon itself.
3. Attendance is limited to approved guests only; please refrain from bringing external guests.
Good luck and have fun!