GPU Monitoring - Hyperbolic Labs

Most On-Demand and Reserved instances ship with GPU monitoring enabled. There is nothing to install and no extra charge — metrics start appearing within a few minutes of the instance becoming ready. A few machine types do not support it yet; see Instances without monitoring below.

Open monitoring for an instance

From your instance’s details page, click Monitoring. The view is scoped to that instance and lists each GPU attached to it.

What you can see

Metric	What it tells you
GPU utilization	Percentage of time each GPU was executing kernels. Sustained low utilization during training usually means a data-loading or CPU bottleneck, not a GPU problem.
GPU memory	Used vs. total per GPU. Running near the ceiling risks out-of-memory errors; see the CUDA out of memory fix.
Temperature	Per-GPU. Sustained readings well above the pack are worth flagging to support — thermal throttling shows up as lower clocks and slower steps before it shows up as an error.

Metrics refresh automatically about every 30 seconds.

Instances without monitoring

Monitoring depends on the platform agent being present on the instance. If an instance’s details page has no Monitoring button, that machine type doesn’t support it yet; nvidia-smi on the instance gives you the same numbers on demand — see Managing Instances.

What monitoring is not

Monitoring is a read-only view. It does not alert you, and it does not restart or replace a GPU automatically. If you see a GPU that has dropped out or is throttling, run the checks in Verifying Instance Performance and contact support with the instance ID and the nvidia-smi -q output.