Editorial illustration for GEAK V3 Boosts AMD GPU Kernels 2.78× with Agent-Driven Optimization
GEAK V3 Boosts AMD GPU Performance 2.78× Faster
AMD's GEAK framework just moved from writing individual GPU kernels to rewriting entire codebases. GEAK V3, detailed in a new blog post from AMD, extends the agent-driven optimization system to work across three kernel languages: HIP, Triton, and FlyDSL. It now runs on both CDNA and RDNA architectures, folding what used to be separate tuning efforts into a single workflow.
The pitch is straightforward: hand-tuning GPU kernels takes real expertise and a lot of manual iteration, and most teams don't have engineers to spare for that work at scale. GEAK V3 tries to close that gap with task planning, automatic test-harness discovery, patch-based handling for kernels spread across multiple files, a dynamic memory system, and a database of expert knowledge the agent can draw on mid-optimization. Instead of just spitting out a kernel, the system now operates at the repository level, producing patches that come with profiling data and test results attached, so an engineer can review the change rather than just trust it.
AMD says the framework is already running in production, including inside an MLPerf inference submission. Below is a look at the specific gains that work produced.
GEAK v3 achieves 3.02× and 2.22× speedups on HIP and Triton benchmarks, respectively, which include kernels spanning a wide range of complexity levels, from individual kernels to full repository-level workloads.
Why this matters
AMD's GEAK v3 numbers matter less as a benchmark flex and more as a signal about where kernel engineering is headed. A 2.78× geomean speedup on L3 kernels, the hardest category in the test set, suggests agent-driven optimization is starting to handle the messy, repository-level work that used to require someone who actually understood HIP or Triton internals. That's the part worth watching for developers: the framework isn't just tuning toy kernels, it's doing task planning and patch generation across a real codebase, which is a different problem than the usual "optimize this one function" demo.
For teams building on AMD hardware, this narrows the tooling gap with CUDA-first shops that have had mature kernel-tuning pipelines for years. For researchers, the GEAK v2-to-v3 comparison on shared Triton kernels gives a rare same-framework benchmark to check whether these agent gains are real progress or just better cherry-picking of test cases. Worth checking AMD's own published methodology before betting a production pipeline on it.
Related Reading
Open Source Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source SDOF Adds Two Defensive Layers via Intent Router and StateAwareDisp
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
From Archives Google Bakes Part of Gemini AI Directly Into "Frozen v2" Chip
From Archives AI Agent Breached Hugging Face as Safety Guardrails Blocked Defenders
Common Questions Answered
What are the three kernel languages that GEAK V3 supports for optimization?
GEAK V3 extends agent-driven optimization to work across HIP, Triton, and FlyDSL kernel languages. This multi-language support allows developers to optimize GPU kernels written in different frameworks within a single unified workflow, eliminating the need for separate tuning efforts for each language.
What speedup improvements does GEAK V3 achieve on HIP and Triton benchmarks?
According to AMD's testing, GEAK V3 achieves 3.02× speedups on HIP benchmarks and 2.22× speedups on Triton benchmarks. These improvements span a wide range of complexity levels, from individual kernels to full repository-level workloads, demonstrating the framework's effectiveness across different optimization scenarios.
How does GEAK V3 differ from previous versions in terms of scope?
GEAK V3 has expanded from writing individual GPU kernels to rewriting entire codebases using agent-driven optimization. This shift represents a significant advancement, as the framework now handles repository-level work and task planning that previously required manual expertise from developers who understood HIP or Triton internals.
Which AMD GPU architectures does GEAK V3 support?
GEAK V3 runs on both CDNA and RDNA AMD GPU architectures, consolidating what used to be separate tuning efforts into a single optimization workflow. This dual-architecture support enables broader compatibility across AMD's GPU product lineup.
What is the geomean speedup that GEAK V3 achieves on L3 kernels?
GEAK V3 achieves a 2.78× geomean speedup on L3 kernels, which are the hardest category in the test set. This significant improvement on complex kernels demonstrates that agent-driven optimization can now handle the messy, real-world repository-level work that traditionally required specialized kernel engineering expertise.
Further Reading
- Benchmarking AI Coding Agents for GPU Kernel Optimization - AMD ROCm Blog
- Introducing Triton Kernel AI Agent & Evaluation Benchmarks - ArXiv
- GEAK: Introducing Triton Kernel AI Agent & Evaluation Benchmarks - AMD ROCm Blog
- GEAK Agent-Driven Optimization of the DeepSeekV4 MLA - AMD ROCm Blog
- GEAK-Triton v2 Family of AI Agents: Kernel Optimization for AMD GPUs - AMD ROCm Blog
