Okay, here is the abstract and summary for the provided transcript.
Abstract:
This presentation announces the AMD Developer Challenge 2025 Imprint Sprint, a kernel optimization competition hosted on the GPU Mode platform in collaboration with the Kernelbot team, Data Monsters, and AMD. Participants are challenged to write highly optimized implementations for three specific kernels relevant to modern AI models (FP8 blockwise gem, Fused MoE, and MLA with Rope) targeting AMD's MI300 GPUs. The competition offers substantial prizes, including up to $100k for achieving near-theoretical performance ("speed of light") on any single kernel and a trip to San Jose for an awards ceremony. Submissions are managed via the Kernelbot job platform, accessible through Discord or a newly introduced CLI tool, with a dedicated website (gpumode.com) providing resources and leaderboards. The presentation details the competition structure, timeline, prize eligibility requirements, submission process improvements, and provides a deep dive into the first kernel challenge: the FP8 blockwise gem, including its specific implementation requirements, memory layout (column-major inputs), and blockwise scaling factors.
0:00:10 Introduction: The talk introduces the AMD Developer Challenge 2025 Imprint Sprint, a collaboration between the Kernelbot team (GPU Mode), AMD, and the Data Monsters team.
0:01:15 Competition Overview (TLDR): This is the first major kernel writing competition round focusing on authoring specific reasoning kernels: FPA gem, MLA (Multi-Head Latent Attention) with Rope, and a Fused MoE kernel.
0:01:39 Grand Prize: Participants can win up to $100k in prizes for achieving performance close to the theoretical maximum ("speed of light"), plus a flight ticket to San Jose for an awards ceremony with AMD.
0:01:52 Platform & Access: The competition is hosted within the GPU Mode server using the Kernelbot job submission platform. Access to MI300 hardware is provided through Kernelbot, making participation possible even without owning the hardware.
0:02:14 Kernels & Resources: Three main kernels are involved. Reference implementations (functional PyTorch, baseline Triton) and target input shapes/configurations will be provided in a repository.
0:02:58 Prize Structure: Top 3 placements (based on aggregate rank across kernels) win up to $25k. Getting close to the roofline on any individual kernel makes one eligible for the larger prize pool (up to $100k).
0:03:18 Key Dates: Submissions open effectively today (April 15th, noon) and run until June 8th, with staggered kernel releases. Pre-registration (required for prizes) deadline is April 30th.
0:03:49 Teams & Eligibility: Teams of up to three people can participate. Anyone can submit to the leaderboard, but registration is mandatory for prize eligibility.
0:04:50 Submission Process (Discord): Submissions use the familiar Kernelbot /leaderboard submit command, targeting MI300 GPUs. Expect fast (~5-10 second) job feedback via GitHub Actions.
0:05:35 New Feature: CLI Submissions: A command-line interface (CLI) tool is being introduced for submitting kernels directly from the terminal, improving developer workflow. (Demo follows).
0:06:56 CLI Demo: The Rust-based CLI allows local testing and leaderboard submissions (currently with a temporary user due to an authentication issue being fixed). Provides results as JSON.
0:09:38 New Feature: Website: A dedicated website gpumode.com is launched, serving as a central hub for leaderboards, kernel descriptions, reference implementations, resources, and lectures.
0:11:18 New Feature: Improved Discord UI: Submissions now feature a cleaner, progressively updated UI with celebratory messages for top placements and results contained within threads.
0:13:26 Debugging Access: Direct SSH access to MI300s is generally not provided. Debugging relies on the fast iteration loop of the Kernelbot job queue. Exception: Very active/high-performing contributors might be granted special access by contacting powderluv (Anush).
0:15:07 Kernel Overview (Alex): The competition focuses on three single-device kernels related to Deepseek V2/V3 models.
0:15:52 Kernel 1: FP8 Blockwise Gem: The first kernel is an FP8 blockwise matrix multiplication (matmult). Detailed specifications (write-up, reference code, math) will be released shortly. This kernel will be the focus for roughly the first two weeks.
0:16:24 Release Schedule: Kernels have staggered releases (FP8 Gem -> Fused MoE -> MLA) but remain open for submission throughout the entire competition once released.
0:17:45 Kernel Specifications: All relevant shapes and data distributions are public for tuning. Scoring uses the geometric mean across test cases. Input data seeds are now private.
0:19:14 GPU Access Update (from AMD): If participants genuinely require dedicated GPU access for debugging, they should inform the organizers, and AMD will try to assist.
Based on Deepseek's deepgem library (but for AMD/HIP).
Task: Compute C = (alpha * A) @ (beta * B)^T, where scaling factors alpha/beta are applied blockwise.
Inputs: A (M x K), B (N x K) in FP8 (E4M3); scaling factors alpha, beta in FP32. Output C is FP32.
Memory Layout: A, B, alpha, beta are provided in column-major order. Output C is row-major. This aids efficient memory access during kernel implementation.
All inputs reside on HBM; shapes are "nice" (divisible, no complex edge cases).
alpha (A's scaler): Shape M x (K/128). One FP32 scalar applies to each row of A, covering blocks of 128 elements along the K dimension.
beta (B's scaler): Shape (N/128) x (K/128). One FP32 scalar applies to each 128x128 block within B.
Crucial Implementation Note: The reference PyTorch code dequantizes before matmul for simplicity. Participants must perform the matrix multiplication in FP8 first (using hardware acceleration like matrix cores/systolic arrays) and then apply scaling factors for performance.
0:31:09 Shapes & Performance Goal: Fixed, public shapes are provided. Info on theoretical limits ("speed of light") will be available to help competitors gauge performance relative to the hardware maximum.
0:31:46 Grand Prize Scoring: Top ranks (#1, #2, #3) decided by aggregate score. The $100k grand prize pool is targeted at submissions achieving near "speed of light" on any single kernel, independent of the overall ranking.
0:33:44 Kernel Complexity: FP8 Gem is likely the simplest. Later kernels (MoE, MLA) involve loading model weights and may offer more optimization scope.
0:34:34 Conclusion: Final reminders about registration deadline (April 30th for prizes), the goal of creating useful kernels, encouragement to participate, using the Discord channel for updates/questions, and team formation (up to 3).