Tensorflow 2.0 未完全使用 RTX2080

Question

我目前正在使用 tensorflow 2.0 测试我的 RTX2080，但我的 RTX2080 与我笔记本电脑中的 GTX1050Ti 一样快。 这是我当前的代码： https://github.com/clementpoiret/IBM-Capstone-CNN/blob/master/capstone.py

我听说将allow_growth设置为 true，我已经完成并解决了 CuDNN 问题。 问题是 GPU 似乎已锁定。

nvidia-smi 在训练期间总是返回相似的值，python 似乎卡在使用 2655Mb：

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.26       Driver Version: 440.26       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce RTX 2080    Off  | 00000000:2D:00.0 Off |                  N/A |
|  0%   55C    P2    60W / 265W |   2912MiB /  7979MiB |     24%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      1026      G   /usr/lib/Xorg                                200MiB |
|    0      6420      G   cinnamon                                      43MiB |
|    0     17187      C   /home/clementpoiret/anaconda3/bin/python    2655MiB |
+-----------------------------------------------------------------------------+

这是我用来设置 gpu 的代码：

def setup_gpus():
    gpus = tf.config.experimental.list_physical_devices('GPU')
    if gpus:
        try:
            # Currently, memory growth needs to be the same across GPUs
            for gpu in gpus:
                tf.config.experimental.set_memory_growth(gpu, True)
                logical_gpus = tf.config.experimental.list_logical_devices(
                    'GPU')
                print(len(gpus), "Physical GPUs,", len(logical_gpus),
                      "Logical GPUs")
        except RuntimeError as e:
            # Memory growth must be set before GPUs have been initialized
            print(e)

我目前的配置是：

System:    Host: Workstation Kernel: 5.3.8-3-MANJARO x86_64 bits: 64 compiler: gcc v: 9.2.0 Console: tty 0 dm: LightDM 1.30.0 
           Distro: Manjaro Linux 
Machine:   Type: Desktop Mobo: Micro-Star model: MPG X570 GAMING EDGE WIFI (MS-7C37) v: 1.0 serial: <filter> 
           UEFI: American Megatrends v: 1.50 date: 10/29/2019 
CPU:       Topology: 12-Core model: AMD Ryzen 9 3900X bits: 64 type: MT MCP arch: Zen L2 cache: 6144 KiB 
           flags: avx avx2 lm nx pae sse sse2 sse3 sse4_1 sse4_2 sse4a ssse3 svm bogomips: 182474 
           Speed: 4155 MHz min/max: 2200/3800 MHz boost: enabled Core speeds (MHz): 1: 3839 2: 2170 3: 3692 4: 2111 5: 3901 
           6: 2094 7: 2069 8: 2176 9: 2177 10: 4361 11: 2191 12: 2200 13: 2169 14: 2200 15: 2179 16: 3821 17: 2183 18: 2190 
           19: 2179 20: 4356 21: 2198 22: 2201 23: 2197 24: 2109 
Graphics:  Device-1: NVIDIA TU104 [GeForce RTX 2080 Rev. A] vendor: Gigabyte driver: nvidia v: 440.26 bus ID: 2d:00.0 
           chip ID: 10de:1e87 
           Display: server: X.org 1.20.5 driver: nvidia tty: 191x17 
           Message: Unable to show advanced data. Required tool glxinfo missing. 
Audio:     Device-1: NVIDIA vendor: Gigabyte driver: snd_hda_intel v: kernel bus ID: 2d:00.1 chip ID: 10de:10f8 
           Device-2: Advanced Micro Devices [AMD] Starship/Matisse HD Audio vendor: Micro-Star MSI driver: snd_hda_intel 
           v: kernel bus ID: 2f:00.4 chip ID: 1022:1487 
           Sound Server: ALSA v: k5.3.8-3-MANJARO 
Network:   Device-1: Realtek RTL8111/8168/8411 PCI Express Gigabit Ethernet vendor: Micro-Star MSI driver: r8169 v: kernel 
           port: d000 bus ID: 27:00.0 chip ID: 10ec:8168 
           IF: enp39s0 state: down mac: <filter> 
           Device-2: Intel Dual Band Wireless-AC 3168NGW [Stone Peak] driver: iwlwifi v: kernel port: d000 bus ID: 29:00.0 
           chip ID: 8086:24fb 
           IF: wlp41s0 state: up mac: <filter> 
Drives:    Local Storage: total: 2.84 TiB used: 90.16 GiB (3.1%) 
           ID-1: /dev/nvme0n1 vendor: Corsair model: Force MP300 size: 111.79 GiB speed: 15.8 Gb/s lanes: 2 serial: <filter> 
           rev: E8FM12.0 scheme: GPT 
           ID-2: /dev/nvme1n1 vendor: Samsung model: SSD 970 EVO 250GB size: 232.89 GiB speed: 31.6 Gb/s lanes: 4 
           serial: <filter> rev: 2B2QEXE7 scheme: GPT 
           ID-3: /dev/sda vendor: Western Digital model: WD30EZRZ-00GXCB0 size: 2.73 TiB speed: 6.0 Gb/s rotation: 5400 rpm 
           serial: <filter> rev: 0A80 scheme: GPT 
Partition: ID-1: / size: 227.74 GiB used: 64.62 GiB (28.4%) fs: ext4 dev: /dev/nvme1n1p2 
Sensors:   System Temperatures: cpu: 67.2 C mobo: 32.0 C gpu: nvidia temp: 53 C 
           Fan Speeds (RPM): fan-1: 0 fan-2: 2017 fan-3: 0 fan-4: 0 fan-5: 0 fan-6: 0 fan-7: 0 gpu: nvidia fan: 30% 
Info:      Processes: 406 Uptime: 4d 14h 26m Memory: 31.37 GiB used: 6.09 GiB (19.4%) Init: systemd v: 242 Compilers: 
           gcc: 9.2.0 Shell: zsh v: 5.7.1 running in: tty 0 inxi: 3.0.36

我听说糟糕的 CNN 性能可能是由于数据存储在 HDD 而不是 SSD 上造成的，但我在两个驱动器上实现了相同的性能。

你有什么主意吗？

谢谢

Answer 1

我也得到了这个——在我的 RTX2080 和我的 GTX960 上做到了。 如果我使用我的 Quadro M2000 或 P4000 卡，它会进入 P0 模式，全速。 NVidia 是否故意限制其非“专业”卡的 TF 性能以促使人们付费？