AMD has launched GEAK v3, an advanced agent-driven framework for optimizing GPU kernels across multiple languages and architectures. The tool is designed to automate the process of kernel optimization, enabling developers to achieve better performance and efficiency in their applications. GEAK v3 supports three kernel languages — HIP, Triton, and FlyDSL — and both CDNA and RDNA GPUs, offering a unified optimization workflow. This framework aims to streamline the optimization process by integrating profiling, testing, and LLM-guided iteration to produce reviewable patches. The results show significant improvements in kernel performance, with notable speedups across a range of workloads. The tool is already being used in production environments, supporting various optimization use cases. Source: amd
GEAK v3 represents an evolution from its previous versions, transitioning from a kernel-generation assistant to a fully automated agent-driven framework. This version introduces enhanced features such as task planning, test-harness discovery, and dynamic memory systems, allowing for more comprehensive optimization. The framework also incorporates an expert knowledge database, which helps guide the optimization process. These improvements have led to measurable gains in performance, with GEAK v3 achieving speedups of 3.02× on HIP benchmarks and 2.22× on Triton benchmarks. The tool's ability to handle complex workloads, from individual kernels to full repository-level tasks, underscores its versatility and effectiveness. Source: amd
The blog post outlines the development and evolution of the GEAK framework, starting with its initial version, which focused on instruction-to-Triton generation. GEAK v1 introduced a modular system with generator, evaluator, reflector, and optimizer modules, enabling iterative code generation and performance tuning. Subsequent versions, such as GEAK v2, expanded the framework into a family of specialized agents, introducing features like multi-offspring evolution and hardware-aware profiling. These enhancements have led to significant improvements in execution accuracy and speedup, with GEAK v2 achieving up to a 9.76% increase in execution accuracy over its predecessor. Source: amd
