New GPU MODE Virtual Hackathon: E2E Model Speedrun

We are excited to announce the launch of the GPU MODE Virtual Hackathon: E2E Model Speedrun, sponsored by AMD. This global competition challenges developers, researchers, and performance engineers to push the limits of large language model (LLM) inference performance on open models optimized for AMD Instinct™ MI355X GPUs.

With a total prize pool of $1.1 million, the hackathon brings together the global developer community to explore advanced GPU kernel optimization and end‑to‑end inference acceleration. Participants will compete to deliver breakthrough performance improvements on real-world LLM workloads while showcasing innovative approaches to GPU performance engineering.

Competition Structure

The hackathon is organized in two phases designed to test both low‑level GPU optimization and full-stack inference performance.

Phase 1: Qualifiers (March 6 – March 30, 2026)

Participants will optimize three critical GPU kernels:

• MXFP4 MoE
• MLA Decode
• MXFP4 GEMM

Teams will be ranked on a public leaderboard based on performance metrics defined in the competition rules. The top 10 teams will advance to the finals.

Phase 2: Finals (March 31 – May 11, 2026)

Finalists will focus on end‑to‑end inference optimization of selected LLM workloads including DeepSeek‑R1 and GPT‑OSS, aiming to achieve breakthrough performance on standardized benchmarks running on AMD Instinct MI355X GPUs. The competition will conclude with an awards ceremony on May 18 recognizing the top teams.

$1.1 Million Prize Pool

Finalists will compete for a share of the $1.1M prize pool across two independent tracks. Finalists may compete in one or both tracks.

Track 1 – DeepSeek‑R1‑0528

• 1st Place: $650,000
• 2nd Place: $30,000
• 3rd Place: $20,000

Track 2 – GPT‑OSS‑120B

• 1st Place: $350,000
• 2nd Place: $30,000
• 3rd Place: $20,000

Who Should Participate

The hackathon is open to eligible individuals or teams of up to three members who are passionate about GPU optimization, AI systems performance engineering, and LLM inference acceleration. Developers, ML researchers, systems engineers, and open‑source contributors are all encouraged to participate.

How to Get Started

Join the virtual hackathon here

Participants should join the AMD AI Developer Program, review the reference kernels for the qualifier stage, and submit optimization results using the Popcorn CLI submission process. Developers can also join the GPU MODE Discord community for discussions, support, and updates throughout the competition. If you are excited about pushing the boundaries of AI performance and competing with some of the best developers in the world, this is your opportunity to participate, learn, and win.