Is there exists some difference between multi-GPU and GPU cluster

Hello, I need some help about GPU cluster. I want to build and GPU cluster enviroment, right now I only have this idea about building GPU cluster, which is with several Nvidia S1070 cards, and S1070 itself is a multi-GPU, am I right? And we can connect the S1070 by infinitband to set up a GPU cluster.

I have two questions here need your help.

The first is “is there exists some difference between multi-GPU and GPU cluster”;
The second is “how to build a GPU cluster enviroment”.

Thanks for your help very much!

Hello,

A multi-gpu cluster is an environment where you have several gpus. By default, you always use one gpu (the device 0). If you want to use another device, you can use the function cudaSetDevice(int device) (0 <= device <= number of devices - 1), it records device as the device on which the active host thread executes the device code.

“How to build a GPU cluster enviroment ?” ==> put several gpus on a same host.

There are some environments like StarPU or GPUSS that allows you to use efficiently and easily a heterogeneous system with several CPUs/GPUs.

Hello,

A multi-gpu cluster is an environment where you have several gpus. By default, you always use one gpu (the device 0). If you want to use another device, you can use the function cudaSetDevice(int device) (0 <= device <= number of devices - 1), it records device as the device on which the active host thread executes the device code.

“How to build a GPU cluster enviroment ?” ==> put several gpus on a same host.

There are some environments like StarPU or GPUSS that allows you to use efficiently and easily a heterogeneous system with several CPUs/GPUs.