Nvidia-smi Memory-Usage of different GPUs always same

I found Memory-Usage of GPU1 is always same as GPU0. Actually they should be different.
What’s the problem?
Thanks a lot.

OS: Windows 11 Pro N 23H2 (22631.3155)
Driver: NVIDIA Studio Driver 551.61

nvidia-smi --version
NVIDIA-SMI version : 551.61
NVML version : 551.61
DRIVER version : 551.61
CUDA Version : 12.4

nvidia-smi nvlink -s
GPU 0: NVIDIA GeForce RTX 3090 (UUID: GPU-3d761d93-e4be-fdaf-f314-b3da9c94b7b8)
Link 0: 14.062 GB/s
Link 1: 14.062 GB/s
Link 2: 14.062 GB/s
Link 3: 14.062 GB/s
GPU 1: NVIDIA GeForce RTX 3090 (UUID: GPU-907b5464-99e5-7ae6-4317-1778bff7d577)
Link 0: 14.062 GB/s
Link 1: 14.062 GB/s
Link 2: 14.062 GB/s
Link 3: 14.062 GB/s

one possible reason is if you have SLI enabled.

If Memory-Usage is affected by NVLink, (consider the memory of two GPUs as a whole), then the displayed Memory-Usage should be the sum of the memory of both GPUs. However, the reality is that the current display is showing the Memory-Usage of GPU0, which is far less than the sum of the memory of the two GPUs.

It isn’t.

If you are asking about Memory usage, i.e. that which is displayed in your black-picture nvidia-smi output, under the middle column “Memory-Usage”, it has nothing to do with nvlink, and I’m not sure why you are involving nvlink in this question, and my comments don’t pertain to nvlink.

I thought “NVLink bridge” and “SLI enabled” were basically the same thing on the RTX 3090, right?

no, they are not

SLI enablement is something you do in the control panel on windows

Try to disable SLI. Now the two Memory-Usage are different.

I’m confusing the two, where can I get more detailed information about them?
If my purpose is to load LLM, do I need to disable SLI?
Thanks a lot.

You don’t need SLI enabled for most/typical “compute” usage of GPUs.

SLI is a mechanism to couple GPUs together to split graphics rendering tasks. You can find plenty of information about SLI with a bit of google searching, and it is briefly covered in the CUDA programming guide.

OTOH, having NVLINK bridge(s) installed can be a very useful feature for compute work on GPUs (and it can also benefit SLI).

Ok. Thanks.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.