ANVILBENCH vikramtiwari/anvilbench
Leaderboard 8 Domain Pillars 100 Tasks & Shaders M4 Max Hardware Methodology & License
View on Kaggle Benchmarks ↗
Kaggle Benchmark v13 Apache-2.0 License
Updated Oct 2026

AnvilBench: Metal Compute Kernel Benchmark

Evaluating frontier LLMs across 100 Metal Shading Language (.metal) compute kernels, 8 Domain Pillars, and 300 hardware verification checks.

Top Model (Kaggle Score) 0.59040 Gemini 4 Argon (59.04% Mean Score)
Suite Coverage 100 Tasks 8 Pillars • 32 Categories • 300 Checks
Closed-Loop Verification 100 / 100 FP64 ≤ 1e-3 + Held-Out ΦT=1 + NaN Guard
M4 Max Hardware Parity ρ = 0.9960 546 GB/s Peak • 3.45× Champion Speedup
AnvilBench — Unified-Memory GPU & Metal Shading Language Compute Kernel Benchmark by Vikram Tiwari. Licensed under Apache-2.0 (benchmarks/LICENSE).
Kaggle Leaderboard Leaderboard 8 Pillars 100 Tasks M4 Max Telemetry Methodology