Nsight Systems Utilization NVLink Metric Appears Incorrect on GB200

Developer Tools Nsight Systems Profiling Linux Targets
1

I am testing NVLink bandwidth on a GB200 using the following command:

nsys profile --gpu-metrics-devices all ./nvbandwidth -t device_to_device_memcpy_write_sm -b 1024

My nvbandwidth tool reports a bandwidth of 717 GB/s (~80% of the theoretical 900 GB/s peak). This result is correct, as it’s based on Data Size / Time Duration. However, the Nsight Systems report shows a utilization of only ~66%.

image
image748×406 74.4 KB

image
image1062×318 80.8 KB

Why is there a difference between these two results? How can I accurately measure the effective NVLink utilization rate with Nsight Systems?

Versions:

  • Driver Version: 570.172.08
  • CUDA Version: 12.9
  • Nsight Systems CLI: 2025.3.1.90-253135822126v0
  • Nsight Systems GUI: 2025.5.1
2

@pkovalenko

3

dcgm monitoring shows the NVLink bandwidth as 700GB, which proves that the output from NVBandwidth is normal, while there is an issue with the output from NSys.

image
image1614×994 99.7 KB

4

ping @pkovalenko

5

Thanks for your feedback. This is a known issue, but currently there’s no immediate solution that I can share. There’s an ongoing internal discussion on how to resolve this.

6

Thanks for your reply! Do you have a rough ETA for the fix (e.g., weeks, months, or longer)?

In the meantime, is there a manual workaround? For instance, can I get the correct value by applying a certain percentage to the User Data and protocol data?

7

gb10x.config.zip (1.9 KB)

No ETA for now, but in the mean time you can try this metric config.

$ nsys profile --gpu-metrics-devices=all --gpu-metrics-set=file:/path/to/gb10x.config

Percentage values are still not right, but it also has bytes/s and GB/s values that should be correct.