← Device AI IQ

GPU Benchmark

GPU Compute Benchmark

1024×1024 tiled matmul · workgroup shared memory · 12s sustained load · FP32 + FP16 · ranked against 24 Apple Silicon chips

🧮

Tiled MatMul

Workgroup Shared Memory tiling for maximum GPU utilization

⚡

FP32 + FP16

FP32 measures general compute, FP16 measures AI inference (~2× faster)

📈

Sustained 12s

12-second load, last 40% averaged for thermal steady-state performance

🔍

Chip Match

Ranked against 24 Apple Silicon chips · FP16 weight 70% · FP32 30%

approx. 30–40 seconds · local only · no data sent · plug in recommended

⚠ If the test fails, make sure hardware acceleration is enabled

🌐Chrome / Edge
  1. Menu → Settings
  2. System → Use hardware acceleration when available
  3. Relaunch browser
  4. Visit chrome://gpu → confirm WebGPU: Enabled
🧭Safari
  1. Safari → Preferences → Advanced
  2. Check "Show Develop menu in menu bar"
  3. Develop → Experimental Features → WebGPU (must be checked)
  4. Requires macOS 13+ / iOS 16.4+

macOS 26 (Tahoe) users: Chrome may blocklist WebGPU on Tahoe due to unrecognized OS version. Fix: ① visit chrome://flags/#enable-unsafe-webgpu → set to Enabled, then relaunch; ② or just use Safari (native WebGPU, zero config).

Firefox does not support WebGPU by default. Chrome 113+ and Edge 113+ have WebGPU enabled out of the box.

AtomGradient · atomgradient.github.io