mosdeo / TensorFlowBenchmark

TensorFlowBenchmark

Docker_arm64v8

• for Apple Silicon M1/M2/...
• 沒有寫 docker-compose.yml，因為找不到使用 metal GPU 的參數
• 目前似乎無法透過 docker 使用 metal GPU
• 只能不用 docker 直接執行 python3 verify.py 來測試

Docker_amd64

• for Intel CPU with CUDA GPU
• 可用 docker run --rm --gpus all -i verify
• 或者 docker-compose up -d 啟動、docker-compose down 停止

參考數據

• MacBook Air M2 8CPU/10GPU 24/2T : 80.4s/epoch (GPU)

Epoch 1/5
782/782 [==============================] - 82s 100ms/step - loss: 4.5111 - accuracy: 0.0912
Epoch 2/5
782/782 [==============================] - 77s 99ms/step - loss: 4.1982 - accuracy: 0.1110
Epoch 3/5
782/782 [==============================] - 78s 99ms/step - loss: 3.6741 - accuracy: 0.1700
Epoch 4/5
782/782 [==============================] - 82s 105ms/step - loss: 3.4374 - accuracy: 0.2094
Epoch 5/5
782/782 [==============================] - 83s 106ms/step - loss: 3.2929 - accuracy: 0.2468

• NVIDIA GA100 [A100 PCIe 40GB]x2 : 47.6s/epoch (Only 1GPU, usage=33%)

782/782 [==============================] - 73s 60ms/step - loss: 4.5955 - accuracy: 0.0829
Epoch 2/5
782/782 [==============================] - 40s 51ms/step - loss: 4.0041 - accuracy: 0.1376
Epoch 3/5
782/782 [==============================] - 40s 51ms/step - loss: 3.9703 - accuracy: 0.1533
Epoch 4/5
782/782 [==============================] - 45s 57ms/step - loss: 4.0180 - accuracy: 0.1366
Epoch 5/5
782/782 [==============================] - 40s 52ms/step - loss: 3.6619 - accuracy: 0.1691