Image Use
Last Updated：2020-09-04
Operating Instruction for GPU Exclusive Image
The exclusive image of GPU integrates GPU drive, CUDA, cuDNN, and mainstream deep learning framework, and other features, the details are as follows:
| GPU Exclusive Image | CUDA Versions | cuDNN | Deep learning framework version |
|---|---|---|---|
| Ubuntu 16.04 LTS amd64 (64bit)-CUDA9.2 | CUDA 9.2 | cuDNN 7.1.4 | No |
| Ubuntu 16.04 LTS amd64 (64bit)-CUDA8.0 | CUDA 8.0 | cuDNN_v5.1, cuDNN_v6 | No |
| 16.04 LTS amd64 (64bit)-CUDA9.0-deep learning development image | CUDA 9.0 | cuDNN 7.1.4 | TensorFlow_gpu 1.10.1、PaddlePaddle_gpu 0.14.0、Caffe2 |
| Ubuntu 16.04 LTS amd64 (64bit)-CUDA8.0-paddlepaddle_0.11.0 | CUDA 8.0 | cuDNN 7.0 | PaddlePaddle_gpu 0.11.0 |
| Ubuntu 16.04 LTS amd64 (64bit)-CUDA8.0-tensorflow_1.3.0 | CUDA 8.0 | cuDNN 7.0 | TensorFlow_gpu 1.3.0 |
| CentOS 7.5 x86_64 (64bit)-CUDA10.0 | CUDA 10.0 | No | No |
| CentOS 7.5 x86_64 (64bit)-CUDA9.2 | CUDA 9.2 | cuDNN 7.1.4 | No |
| CentOS 6.8 x86_64 (64bit)-CUDA9.2 | CUDA 9.2 | cuDNN 7.1.4 | No |
Operating Instruction:
- If you purchase NVIDIA Tesia V100 GPU card, please select CUDA9.2 version image to create the GPU instance. With CUDA 8.0 version image, V100 GPU card cannot be identified in the GPU instance.
- When you use 16.04 LTS amd64 (64bit)-CUDA9.0-deep learning development image, multiple pre-installed frameworks are available. You can use the frameworks by sandbox of source, for example, switch to Caffe 2 environment with the following mode: # source /home/Caffe2-py27/bin/activate.
- As cuDNN has been downloaded to the directory /home/work/cudnn/, you can install by manual the Deep Learning framework if the deep learning framework is not installed in the GPU exclusive image. For more information, see the official documents of NVIDIA:https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html. You can also download by yourself if you need other version cuDNN:https://developer.nvidia.com/rdp/cudnn-download.
Steps to Use Public Image
BCC public images don't include the officially published version of GPU driver. You can install NVIDIA GPU driver by reference to the following mode:
The download of GPU driver:
- Download the driver program mapped with the type of GPU in instance, from the NVIDIA official website. Address:http://www.nvidia.com/Download/index.aspx?lang=cn.
-
Find by manual the driver program applicable to the GPU instance with attention paid to the type of GPU and operating system version:
- Click the search button and enter into the download page.
-
After confirming the information, click the download to download the driver program to local.
-
Upload the driver program onto CVM server and install it according to the instructions on the official website of NVIDIA:
- From December 19th, 2017, the support of GPU instance for Windows is suspended. You cannot select the Windows image to create new GPU instance. For the follow-up time to provide Windows image support, please pay attention to the official announcement.