I know SLI for the current 1080Ti only supports 2 ways. Does that means even though I put 3 or 4 1080Ti on 4-ways SLI supported Motherboards, the compiler will only access two graphic cards? Moreover how will SLI function with 4 1080ti graphic cards under these two conditions?
1. Run Tensorflow-gpu (only one graphic card work? or a pair of 1080Ti?)
2. program with cuda (Does the compiler simply treat these 4 graphic cards independently? There is nothing related to SLI?)
SLI doesn’t have anything directly to do with CUDA. From a programming model perspective, they are mostly orthogonal.
For CUDA, the best SLI setup is to completely disconnect your SLI bridges and remove them. Disable SLI.
“First, an allocation in one CUDA device on one GPU will consume memory on other GPUs that are part of the SLI configuration of the Direct3D or OpenGL device. Because of this, allocations may fail earlier than otherwise expected.”
Is that the benavior you want? Probably it isn’t. So disable SLI.
On linux, for best compute performance, it’s also recommended that you remove your GPUs from any X display usage/configuration.
SLI is about fast graphics, not compute.
Thank you for the clear explanation. It is quite helpful. I know what to do now.