Two GPU cards on a single machine

Accelerated Computing CUDA CUDA Programming and Performance
1

I am testing deep learning on multiple GPU cards using TensorFlow. The performance of single GPU is better compared to 2 GPUs.

Following is the result of simpleP2P utility,

/usr/local/cuda/samples/0_Simple/simpleP2P# ./simpleP2P
[./simpleP2P] - Starting…
Checking for multiple GPUs…
CUDA-capable device count: 2

GPU0 = “GeForce GTX 750 Ti” IS capable of Peer-to-Peer (P2P)
GPU1 = “GeForce GTX 750 Ti” IS capable of Peer-to-Peer (P2P)

Checking GPU(s) for support of peer to peer memory access…

Peer access from GeForce GTX 750 Ti (GPU0) → GeForce GTX 750 Ti (GPU1) : No
Peer access from GeForce GTX 750 Ti (GPU1) → GeForce GTX 750 Ti (GPU0) : No
Two or more GPUs with SM 2.0 or higher capability are required for ./simpleP2P.
Peer to Peer access is not available amongst GPUs in the system, waiving test.

Following is the result of simpleMultiGPU utility,

/usr/local/cuda/samples/0_Simple/simpleMultiGPU# ./simpleMultiGPU
Starting simpleMultiGPU
CUDA-capable device count: 2
Generating input data…

Computing with 2 GPUs…
GPU Processing time: 13.510000 (ms)

Computing with Host CPU…

Comparing GPU and Host CPU results…
GPU sum: 16777280.000000
CPU sum: 16777294.395033
Relative difference: 8.580068E-07

Could anyone suggest how to enable Peer to Peer memory access?

2

Peer to peer (P2P) also depends on what type of system the GPUs are plugged into.

That is most likely the issue in your case. You’ll find quite a few discussions around enabling P2P on this forum if you want to search for them.

3

The operating system is Ubuntu 14.04. Is this what you mean by type of system?
Could you please share any post regarding it?

4

run this search through google:

site:devtalk.nvidia.com peer topology

and start reading.