Speedup TensorFlow with GPU

[jinzishuai]

Examples

Course 4, Week 2 HW

https://github.com/jinzishuai/learn2deeplearn/blob/master/deeplearning.ai/C4.CNN/week2_DeepModelCaseStudy/hw/ResNets/ResidualNetworks.py

It takes about 5 minutes to finish only 1 epochs on Coursera.

On Ubuntu-16 Virtualbox VM on PC

[jinzishuai]

Mac OS X without GPU

Much faster (at least 3 times)

[jinzishuai]

Ubuntu laptop without GPU

• Epoch 1: 94s 87ms/step
• Epoch 2: 92s 85ms/step

[jinzishuai]

Windows with GPU but out of memory on the same laptop

We have 2GB of GPU Memory

C:\Program Files\NVIDIA Corporation\NVSMI>nvidia-smi.exe
Sat Nov 18 19:08:12 2017
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 376.51                 Driver Version: 376.51                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro K1100M      WDDM  | 0000:02:00.0     Off |                  N/A |
| N/A   46C    P8    N/A /  N/A |     28MiB /  2048MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

C:\Program Files\NVIDIA Corporation\NVSMI>

[jinzishuai]

EC2 GPU Instance

http://docs.aws.amazon.com/dlami/latest/devguide/tutorial-conda.html#tutorial-conda-switch-tf

• to use tensorflow: source activate tensorflow_p27 (note that there is some problem with tensorflow_p36 due to https://github.com/tensorflow/tensorflow/issues/14182)

First Run

[jinzishuai]

Second EC2 GPU Run

It seems that with GPU, the first run takes longer time to initialize while the following runs are much faster.

also notice that there is 11GB of GPU memory.

[jinzishuai]

More information about p2 instances

https://aws.amazon.com/ec2/instance-types/p2/

[jinzishuai]

https://aws.amazon.com/ec2/instance-types/#p2

Use Cases: Machine learning, high performance databases, computational fluid dynamics, computational finance, seismic analysis, molecular modeling, genomics, rendering, and other server-side GPU compute workloads.

[jinzishuai]

Conclusion

The Linux tf-1.5 build does not support my GPU but the Windows build does.

So I would always run my tensorflow code on Windows.