AKG Kernel Agent Automates Kernel Generation for AI Workloads

This paper addresses the critical bottleneck of manual kernel optimization in AI system development, particularly given the increasing complexity of AI models and the diversity of hardware platforms. The proposed multi-agent system, AKG kernel agent, leverages LLM code generation to automate kernel generation, migration, and tuning across multiple DSLs and hardware backends. The demonstrated speedup over baseline implementations highlights the practical impact of this approach.

• •Addresses the kernel optimization bottleneck in AI.
• •Proposes a multi-agent system (AKG kernel agent) for automated kernel generation.
• •Supports multiple DSLs and hardware backends.
• •Demonstrates performance improvements over baseline implementations.