Researchers from Harvard, Stanford, Together AI, and Caltech have released Hawkeye, an open-source framework designed to help autonomous coding agents produce GPU kernels that are optimized for specific hardware. The project set out to answer how coding agents can be made hardware-aware with only minimal expert intervention.
What Hawkeye does
Hawkeye organizes kernel generation around a compact, comprehensive taxonomy of unit tests. The framework packages curated information about different hardware platforms and corresponding optimization strategies into unit tests that an agent can read, invoke, or use as code fragments to compose larger kernels.
Each unit test pairs a human-authored solution kernel with a profiling metric that verifies whether the optimization succeeded. The solution kernel is provided as a callable function with a short usage guide so the agent can use it as a syntax example, call it directly, or recompose parts into a new kernel.
Hardware and numeric formats evaluated
The authors evaluated Hawkeye on multiple GPU architectures and numeric precisions: NVIDIA Ampere, NVIDIA Hopper, NVIDIA Blackwell, and AMD MI350, and across BF16, FP8, NVFP4, and MXFP4 precisions.
Key results
On established workloads where torch.compile typically dispatches to expert-tuned vendor libraries such as cuBLAS, cuDNN, and FlashAttention, Hawkeye matched or exceeded those implementations in BF16 and low-precision formats, and in some formats that PyTorch cannot natively run.
For emerging attention variants that torch.compile cannot fuse because they use non-standard scans and gates, Hawkeye produced a substantial improvement: an 18.9× geomean speedup compared to expert-authored Triton kernels from the Flash Linear Attention library in the reported experiments. Across Linear Attention tasks, Hawkeye approaches or outperforms Flash Linear Attention on every tested architecture, with examples including a 1.22× speedup on Blackwell and parity (1.00×) on MI350.
The researchers also report that increasing test-time compute when using Hawkeye yields the most performant kernels across architectures.
Why this matters
The Hawkeye work illustrates that with a relatively small set of curated, high-quality examples and tests, AI systems can learn to reproduce and in some cases surpass hand-optimized, expert human work such as GPU kernels. The implication is that building similar gold-standard helper systems can enable machines to bootstrap beyond human-crafted performance on specific engineering tasks.
Source
Details and experimental results are available in the paper “Hawkeye: Hardware-Aware GPU Kernel Optimization with Minimal Supervision” (alphaxiv).
