# Tesla T4 Benchmark Test
###### tags: `aetherAI`
1. 安裝NVIDIA Tesla卡片驅動 (for Ubuntu 16.04)。

    # 下載驅動
    wget -O nv_driver.deb http://tw.download.nvidia.com/tesla/418.67/nvidia-diag-driver-local-repo-ubuntu1604-418.67_1.0-1_amd64.deb && \
    # 安裝驅動
    sudo dpkg -i nv_driver.deb && \
    sudo apt-get update && \
    sudo apt-get install cuda-drivers && \
    # 重新啟動
    sudo reboot

安裝完成並且重開機後, 於終端機輸入```nvidia-smi```, 應即可查看顯示卡狀態。
2. 參考 [NVIDIA Docker安裝說明](https://github.com/NVIDIA/nvidia-docker#quickstart) 來安裝Docker & NVIDIA-Docker。
3. 跑Benchmark:

    # 下載用來跑Benchmark的Docker容器映像檔
    nvidia-docker pull aetherai/research:cu10.0-dnn7.6-gpu-tf-cv-19.07
    
    # 運行測試範例
    NV_GPU=0 nvidia-docker run -it --rm \
        aetherai/research:cu10.0-dnn7.6-gpu-tf-cv-19.07 python3 /workspace/tf_benchmarks/scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py \
        --num_gpus=1 \
        --model=resnet50 \
        --batch_size=256 \
        --nodistortions \
        --use_fp16 \
        --num_epochs=1

此測試將輸出類似以下的結果:

    ...
    ...
    ...
        Running warm up                                                       [406/1171]
    2019-07-19 04:09:58.783779: I tensorflow/stream_executor/platform/default/dso_lo
    ader.cc:42] Successfully opened dynamic library libcublas.so.10.0
    2019-07-19 04:09:59.197938: I tensorflow/stream_executor/platform/default/dso_lo
    ader.cc:42] Successfully opened dynamic library libcudnn.so.7
    Done warm up
    Step    Img/sec total_loss
    1       images/sec: 651.2 +/- 0.0 (jitter = 0.0)        7.808
    10      images/sec: 652.3 +/- 0.4 (jitter = 1.2)        7.883
    20      images/sec: 652.1 +/- 0.3 (jitter = 1.2)        8.018
    30      images/sec: 651.6 +/- 0.3 (jitter = 1.6)        7.836
    40      images/sec: 651.3 +/- 0.2 (jitter = 1.4)        7.782
    50      images/sec: 651.3 +/- 0.2 (jitter = 1.2)        7.864
    60      images/sec: 651.1 +/- 0.2 (jitter = 1.3)        7.880
    70      images/sec: 651.0 +/- 0.2 (jitter = 1.3)        7.842
    80      images/sec: 650.9 +/- 0.2 (jitter = 1.5)        7.843
    90      images/sec: 650.9 +/- 0.1 (jitter = 1.4)        7.850
    100     images/sec: 650.8 +/- 0.1 (jitter = 1.4)        7.751
    110     images/sec: 650.6 +/- 0.1 (jitter = 1.5)        7.911
    120     images/sec: 650.2 +/- 0.2 (jitter = 1.6)        7.803
    130     images/sec: 650.0 +/- 0.2 (jitter = 1.7)        7.811
    140     images/sec: 650.0 +/- 0.2 (jitter = 1.5)        7.764
    150     images/sec: 649.8 +/- 0.2 (jitter = 1.7)        7.792
    160     images/sec: 649.6 +/- 0.2 (jitter = 1.9)        7.655
    170     images/sec: 649.4 +/- 0.2 (jitter = 2.1)        7.672
    180     images/sec: 649.2 +/- 0.2 (jitter = 2.2)        7.656
    190     images/sec: 649.1 +/- 0.2 (jitter = 2.3)        7.699
    200     images/sec: 649.0 +/- 0.2 (jitter = 2.3)       
    ...
    ...
    ...

4. 於運行Benchmark的同時，另開一個終端機，並且監控卡片溫度:

    chweng@HonghuA:~$ watch -n 1 nvidia-smi
    Every 1.0s: nvidia-smi                             

    Fri Jul 19 12:13:57 2019
    +-----------------------------------------------------------------------------+
    | NVIDIA-SMI 415.27       Driver Version: 415.27       CUDA Version: 10.0     |
    |-------------------------------+----------------------+----------------------+
    | GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
    | Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
    |===============================+======================+======================|
    |   0  TITAN RTX           On   | 00000000:04:00.0 Off |                  N/A |
    | 65%   86C    P2   257W / 280W |  23347MiB / 24190MiB |     98%      Default |
    +-------------------------------+----------------------+----------------------+

可見此卡片(Titan RTX)已達到86度。建議將Tesla T4溫度控制至最高80度上下，以免溫度過高而造成卡片降速。