AMD Developer ChallEnge 2025

Distributed Inference Kernels

Introduction

In this challenge sponsored by Advanced Micro Devices, Inc. (“AMD”), participants are invited to form teams of up to three members. The goal is to develop and optimize multi-GPU communication kernels and achieve significant performance gains in LLM models, ultimately pushing the limits of inference performance on AMD Instinct™ GPUs.

The registration period is from August 23 to September 20, 2025. Kernel submissions will be accepted from August 30 and October 13, 2025.

Participants are eligible to win prizes up to $100,000 and may be selected to travel to San Francisco, CA, for a special awards ceremony.

AMD Developer challEnge

The Challenge Overview

This challenge involves optimization and performance testing of the following multi-GPU communication kernels:

Challenge Overview

KEY KERNELS

The following reference information will be provided at the time of each kernel problem release:

Kernel
Submission Window
Single node 8 GPU all-to-all kernel
August 30 to October 13, 2025
Single node 8 GPU GEMM + Reducescatter kernel
September 13 to October 13, 2025
Single node 8 GPU Allgather + GEMM kernel
September 27 to October 13, 2025
  • Reference implementation (baseline reference)
  • Kernel level roofline performance (theoretical maximum)
  • Target input shapes and configurations
$100,000
Grand prize
$25,000
First Place
$15,000
Second Place
$10,000
Third Place

Prizes

In addition to the chance to win a prize, select participants may be invited to San Francisco, California, USA, to participate in AMD DevDay 2025 during the Open-Source AI Week. Details will be provided to select participants in a reasonable amount of time before the award ceremony, including information regarding travel, hotel accommodations, and any applicable restrictions.

Note

1. Attending the award ceremony on October 20, 2025, is not a requirement for winning a prize.

2. Only kernels from registered participants within the registration deadline will be considered for scoring.

3. Reference kernels and associated information will be provided at the start of the particular schedule.

4. Teams or individuals deemed winners are eligible to receive one (1) prize only, to be awarded at the sole discretion of AMD.

5. Winning a prize depends on participants’ ability to meet specific performance targets. If no team meets the performance targets, AMD reserves the right to refrain from awarding a prize.

6. A participant or team’s place on the leaderboard is not necessarily reflective of being determined to be a winner, nor is leaderboard placement a guarantee of winning a prize.
Four (4) winning submissions will be selected by AMD in its sole discretion after reproduction and verification of results.

7. The challenge will be hosted on the GPU Mode Discord server, operating on their KernelBot automation platform to execute kernel submissions, evaluate and debug performance and populate the results on the leaderboard.

8. The server also enables communication among participants, sharing learning resources, and viewing competition statistics.

Challenge Schedule

August 23, 12:00 PM PST

Registration Open

AUGUST 30, 12:00 PM PST

Kernel Submission Open

SEPTEMBER 20, 11:59 PM PST

Registration Deadline

October 13, 11:59 PM PST

Kernel Submission Deadline

October 20, Details to follow

Awards Сeremony

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.

Challenge policy