WebGPUベンチマーク
ブラウザ上でGPUの計算、メモリ、AI性能を測定し、YOLO推定に利用できます。
実GPUベンチマーク
各処理を自動調整し、中央値でGPU性能を測定します。
YOLO学習設定
YOLO学習設定
YOLO11 Small · 640 · 物体検出
9.4M params · 21.5B FLOPs @ 640
VRAM and speed estimates change with the selected YOLO model and image size. Dataset Images and Epochs are used to estimate training time.
実際のGPUに近いモデルを選択してください。
Compute
Repeated FMA shader throughput
Memory
Storage-buffer read + write throughput proxy
AI Matrix
1024×1024 tiled matrix multiplication
Production use requires HTTPS. localhost can be used for local WebGPU development.
WebGPU Benchmark
Starting the test runs real Compute / Memory / Matrix workloads on this PC's GPU.
How v2 measures
To prevent command-submit overhead from dominating fast GPUs, each workload automatically increases its repeat count. GPU timestamps are preferred when supported; otherwise v2 uses a sufficiently long calibrated wall-clock workload.
Next step
The next step is to calibrate this v2 score against measured RTX 5060 CUDA/PyTorch/YOLO results, enabling WebGPU score → YOLO recommended batch and training-speed estimates.
Important
WebGPU is not the same API or workload as PyTorch CUDA training. At this stage, WebGPU results alone are not used to claim an exact YOLO batch size or training time.
使い方
- 「yolo-gpu-tester」で使う値を入力します。
- オプションを確認して実行します。
- 結果を確認してコピーまたは保存します。