Hawk Boosts NPU Kernel Generation with Hardware-Aware Knowledge

Developing high-performance kernels for Neural Processing Units (NPUs) is a major bottleneck in the industry, often requiring manual navigation of complex hardware constraints and memory hierarchies. While Large Language Models (LLMs) offer automation potential, they typically fail with NPUs due to their lack of inherent hardware-specific knowledge, leading to runtime crashes and performance issues even if code compiles. To overcome this, a new training-free framework called Hawk has been introduced. Hawk integrates hardware-aware knowledge through three core modules. First, a Run-Time Knowledge Synthesis Module uses a Triple-Part Executable Knowledge Representation to link error contexts with executable semantics. Second, a Bottleneck-Aware Knowledge Retrieval Module employs a 2D-Retrieval paradigm, projecting queries into both syntactic and hardware-aligned semantic spaces. Finally, an Effect-Driven Knowledge Distillation Module leverages LLM-driven semantic arbitration to continuously refine knowledge by pruning errors and consolidating redundancies based on empirical execution feedback. Evaluations on real-world NPU workloads demonstrate that Hawk significantly increases generation accuracy from 49.4% to 80.0% and achieves up to a 2.2x execution speedup over current baselines.