proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersMingzhe Du, Anh Tuan Luu, Dong Huang1 min readpaperadvanced

Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles

Summary

The paper introduces mutation analysis as a quantitative adequacy metric for GPU‑kernel benchmark oracles. By injecting 10,303 deterministic faults into 188 verified CUDA kernels (7,384 with a known kill witness), they show the official KernelBench checker misses 16.9% of faults—especially 78.6% of precision‑related faults. Their analysis quantifies the impact of existing patches (e.g., KernelBen…

  • Mutation analysis provides a concrete, repeatable metric for oracle adequacy in GPU kernel benchmarking.
  • The official KernelBench oracle is systematically weak: it fails to detect 1 in 6 witnessed faults, with precision faults being especially problematic.
  • Existing patches can be quantitatively decomposed: hidden inputs contribute +4.0 points, tighter tolerances +4.5 points to the detection score.
  • A published fuzzing recipe is over‑aggressive, rejecting correct kernels 107 times.

Benchmark oracles drive leaderboard rankings and reinforcement‑learning reward signals for LLM‑generated GPU kernels. If the oracle is weak, it can mislead research directions, inflate performance claims, and waste engineering effort on spurious optimizations. A measurable adequacy metric lets the…

8/10

Related reading

  1. Nvidia announces native GPU programming in Rust

    NVIDIA released CUDA‑Rust, letting you write GPU kernels directly in Rust and compile to PTX. Two programming models are supported: the traditional SIMT model via the `cuda-oxide` backend (nightly Rust, custom codegen) and the newer Tile model via `cutile‑rs` (stable Rust, JIT‑compiled Tile IR). Both provide Rust‑typed safety guarantees (e.g., `DisjointSlice`, tensor partitioning) and simple Carg…

    Hacker News front pagenvidia.com11 minHN961402
  2. Accurate Models of AMD Matrix Cores

    The authors reverse‑engineer AMD’s CDNA matrix cores, build per‑architecture MATLAB models that match hardware bit‑for‑bit on 10 M random tests, and use them to compare AMD vs NVIDIA tensor‑core accuracy.

    Hacker News front pagearxiv.org2 minpaperHN7911
  3. NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut

    NVIDIA’s Vera Rubin NVL72 AI inference system shows up to 3.7× higher throughput than the prior GB300 NVL72 on MLPerf v6.1 benchmarks (Qwen3‑VL, DeepSeek‑R1), achieves 99% scaling efficiency across 288 GPUs, and benefits from software optimizations (NVFP4 precision, kernel fusion, disaggregated serving). The post is a product announcement with concrete benchmark numbers but limited technical dept…

    Nvidianvidia.com4 min