Dear @nvidia — nvidia-smi is broken on the DGX Spark

@nvidia — nvidia-smi doesn’t report memory usage on the DGX Spark. I get that the memory is unified, but the lack of an official CLI for inspecting it is a real gap. Please fix ASAP

It’s not exclusively GPU memory. You need to use a different means to query that information.

nv-monitor and dgxtop are my current favorites, open source and built by the community.

Try this one Sparkview — GPU monitor tool with GB10-aware unified memory handling

there are plenty of third-party tools, but when we pay $$$ for hardware, we expect proper support and tooling.

Thanks for the shout out! nvidia-smi is our daily use tool, and it miss this, this is so annoying.

we also use nvidia-smi during training to monitor gpu usage, etc, and need this to work.

Yes… but honestly I trust the community more than the big brands nowadays :) We could either wait, complain and scream or get help by our friends in this forums.

Screaming and then using another tool is perfectly fine though :)

Please reference DGX Spark / GB10 FAQ

@NVES not sure what your comment is in reference too. We are using the dgx-spark for training, and have no proper way to monitor gpu-ram in our terminal.

Apologies, linked to the wrong FAQ section. DGX Spark / GB10 FAQ

The DGX Spark uses a unified memory architecture, so there isn’t dedicated GPU VRAM for nvidia-smi to report on.

To monitor memory usage, you can instead use standard system tools like:
• top
• htop
• free

or check the DGX Dashboard at:
https://localhost:11000

These will give you a more accurate view of actual memory consumption on DGX Spark.

I think this is what is meant