An agent-driven CUDA kernel optimizer that uses LangGraph and a large language model to iteratively generate, validate, benchmark, and refine GPU kernels and launch configurations. It compiles candidate kernels on the fly with NVRTC and launches them via the CUDA Driver API from a C++ harness, while Python orchestrates output comparison (against a NumPy reference), candidate selection, and result recording. The agent can query GPU properties, inspect Nsight Compute counters, and retrieve NVIDIA documentation to guide optimizations; every experiment is recorded and the fastest validated implementation is preserved. Ranking is based on the geometric mean of latency across performance cases, and the model can repair invalid candidates within an iteration budget.
Practical details include development on Windows with an RTX 3060, requirements of Python 3.12+, an NVIDIA GPU and compatible CUDA toolkit/driver, CMake, a C++17 compiler, and an OpenAI API key; the default LLM is gpt-5-mini. Timing defaults to 10 warmups and 100 measured launches; compilation and profiler replay are excluded from ranking. Runs produce a results/run-NNN directory with sources, inputs/outputs, history.json, summary.json, best.cu, and heatmaps; optional flags enable Nsight profiling and NVIDIA research. This is experimental: per-case passes do not guarantee general correctness, generated references are not independent oracles, improvements are workload-dependent, no vendor-library comparisons are included, and generated code/scripts execute locally without sandboxing.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Go 1.26 and 1.27 introduced experimental APIs for SIMD operations across multiple architectures, enabling faster data processing. These APIs replace the previous need for assembly code, improving performance in tasks like cryptography and AI. (go.dev)
Topcoat is a comprehensive full-stack Rust framework designed to simplify building web applications with features like views, components, mailers, and an ORM. It aims to make Rust development as productive and enjoyable as Ruby on Rails, especially for AI-driven server applications. (tokio.rs)
Jev Code Reviewer is an open-source tool designed to assist developers in reviewing code quality and security. It integrates with GitHub workflows to streamline the code review process. (github.com)
Plan modes are becoming obsolete as AI models improve, making detailed instructions less necessary. Maintaining a clear mental model of complex, rapidly changing code is more important than ever to ensure understanding and effective debugging. (aymannadeem.com)
Go versions 1.26 and 1.27 introduced experimental platform-independent APIs for SIMD operations, enabling faster data processing across CPUs. These APIs allow Go programs to utilize SIMD without writing assembly code, improving performance for tasks like cryptography and AI. (go.dev)
Rob Race argues that traditional programming tutorials are becoming obsolete because static, step-by-step instructions no longer meet developers' needs. Instead, he advocates for more dynamic, context-rich educational approaches that focus on real-world application and understanding. (robrace.dev)
Today's best Hacker News stories, summarized and screenshotted, one email a day.