AMD: GEAK v3 Open-Source Agent Speeds Up GPU Kernels 3.02× on Instinct MI300X, MI355X, and RDNA4

What is GEAK v3 and how does it differ from its predecessor?

GEAK v3 is AMD’s open-source agent (Apache 2.0 license, version AMD-AGI/GEAK v3.2.2) for automated GPU kernel optimization — low-level programs that execute parallel computations directly on the graphics processor. Unlike the previous version, which optimized each kernel in isolation, GEAK v3 moves to a repository-level approach: the agent analyzes the entire code repository and optimizes kernels written in HIP, Triton, and FlyDSL in the context of their actual use within the project.

3.02× on HIP and 2.22× on Triton kernels

On 16 HIP kernels, GEAK v3’s agent-driven optimization achieves a geomean speedup — the geometric mean of speedup across the whole set of tested kernels — of 3.02×, and 2.22× on 17 Triton kernels. A direct comparison with the previous version exists on a shared subset of Triton kernels supported by both generations: there, GEAK v3 reaches 1.95× versus just 1.03× for GEAK v2, a practically negligible improvement over unmodified code. The gap shows how much repository-level context helps the agent find optimizations that isolated analysis of a single kernel cannot uncover.

Production evidence on MI300X, MI355X, and RDNA4

AMD cites concrete production examples: the attention kernel of the GPT-OSS 120B model is accelerated 8-10%, while the MLA (multi-head latent attention) kernel of DeepSeek V4 on the Instinct MI355X achieves 2.10× higher throughput and a 3.71× shorter time-to-first-token — the time to the first generated token. A 3D rendering kernel is accelerated 36% while preserving computational accuracy. Testing also covers the RDNA4 architecture, the gfx1201 chip also known as Navi48, extending the tool’s reach beyond data centers to professional graphics cards.

The results show that agent-driven optimization is not limited to the largest Instinct accelerators built for training large models, but also works on the consumer RDNA4 architecture that powers gaming and professional workstations. GEAK v3’s availability under the Apache 2.0 license means development teams outside AMD can take the agent, adapt it to their own repository, and repeat the optimization process on their own kernels without additional licensing. This opens the door to faster iteration on custom GPU code within the community developing HIP, Triton, and FlyDSL projects.