GEAK-Triton v2 Family of AI Agents: Kernel Optimization for AMD Instinct GPUs

2025-12-23 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

AMD has introduced the GEAK-Triton v2 family of AI agents, GEAK-OptimAgentv2 and GEAK-OpenEvolve, designed to automate and optimize GPU kernel generation and refinement for AMD Instinct GPUs. GEAK-OptimAgentv2, an advanced AI agent for Instruction-to-Triton kernel generation, features multi-offspring evolution, an LLM-based evaluator, and a hardware-aware feedback loop, achieving up to a +9.76% accuracy jump and an average 3.32x speedup over reference kernels. GEAK-OpenEvolve is a new Triton-to-Triton framework that uses Quality-Diversity search to optimize existing kernels, demonstrating average speedups of 3.42x on TritonBench-modified and 7.02x on ROCm-bench. These agents leverage AI to address the complexity of manual kernel tuning, enhancing the efficiency of AI model training and inference on AMD hardware.

Key takeaway

For AI Scientists developing or optimizing models on AMD Instinct GPUs, exploring the GEAK-Triton v2 family can drastically reduce manual tuning efforts and improve performance. You should consider integrating GEAK-OptimAgentv2 for generating new, highly efficient Triton kernels and GEAK-OpenEvolve for optimizing existing ones, especially for critical components like LLaMA feedforward or RoPE kernels. Leveraging the hardware-aware feedback loop is crucial for achieving substantial speedups and overcoming performance bottlenecks.

Key insights

AI agents can automate and significantly optimize GPU kernel generation and refinement for AMD Instinct GPUs.

Principles

• Hardware-aware feedback improves kernel optimization.
• Evolutionary search enhances kernel diversity and quality.
• Multi-offspring generation boosts code correctness.

Method

GEAK-OptimAgentv2 uses multi-offspring evolution, an LLM-based evaluator, and a Profiler-Analyzer hardware feedback loop for instruction-to-Triton kernel generation. GEAK-OpenEvolve employs a Quality-Diversity search with MAP-Elites for Triton-to-Triton kernel optimization.

In practice

• Utilize GEAK-OptimAgentv2 for new Triton kernel generation.
• Apply GEAK-OpenEvolve to optimize existing Triton kernels.
• Integrate hardware profiling for targeted kernel improvements.

Topics

• AI Agents
• GPU Kernel Optimization
• Triton Language
• AMD Instinct GPUs
• Evolutionary Algorithms

Code references

• AMD-AGI/GEAK-agent
• AMD-AGI/GEAK-eval
• ROCm/rocm-systems
• unslothai/unsloth

Best for: AI Scientist, AI Engineer, Machine Learning Engineer, Research Scientist

Related on AIssential

• CuTeGen: An LLM-Based Agentic Framework for Generation and Optimization of High-Performance GPU Kernels using CuTe — CuTeGen iteratively refines GPU kernels using CuTe abstraction and delayed profiling, achieving competitive performance.
• ORBIT-2 based Weather and Climate Downscaling and Downscaled Global Forecasts on AMD Instinct — ORBIT-2 downscales global weather data to high resolution efficiently on AMD GPUs using a novel transformer architecture.
• Speeding up GPU kernels by 38% with a multi-agent system — Multi-agent systems can autonomously optimize complex software, achieving significant performance gains in weeks.
• AutoMegaKernel: A Statically-Checked Agent Harness for Self-Retargeting Megakernel Synthesis — AutoMegaKernel synthesizes safe, efficient, and retargetable megakernels for Llama models, outperforming cuBLAS on inference-class GPUs.
• Streamlining Recommendation Model Training on AMD Instinct™ GPUs — ROCm training docker simplifies recommendation model training on AMD Instinct GPUs by optimizing sparse embedding handling.
• Stop hand-tuning kernels: How Neuron Agentic Development accelerates AWS Trainium optimizations — AI agents can automate complex hardware-aware kernel optimization, making performance engineering accessible to more ML developers.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.