Live GPU Metrics
See whether the GPU you are paying for is actually working. Utilization, VRAM, temperature and power on a running deployment, refreshed every few seconds. And a reading the card cannot give you says so, instead of showing a zero.
Provider coverage
Metrics are wired into every deployment on every provider. The agent ships with the box and the Metrics option is on by default. What varies is what the card itself exposes: utilization, memory and clock speeds are read across NVIDIA supply, while temperature and power depend on the GPU's virtualization profile. A vGPU-based card reports neither, so the console shows those two as Unavailable rather than inventing a zero. AMD supply is collected through ROCm-SMI instead of NVIDIA's DCGM.
What you see
Open the Metrics tab on a running deployment and you get, for every GPU on the box: utilization, memory used against total, temperature, power draw, and SM and memory clock speeds. Underneath, the same view covers the host (CPU utilization and system memory) because a GPU sitting idle while the CPU pins at 100% is a data-loading problem, not a GPU problem, and you cannot tell those apart from one number.
Each reading appears as a current value and as a chart over the life of the session, so you can see a training loop settle into a rhythm, or spot the moment throughput fell off.
Why it matters on rented hardware
On your own workstation an underused GPU is a waste of electricity. On a rented one it is a waste of money, billed by the second, and it is invisible unless something tells you. The most expensive failure mode is not a crash: a crash you notice. It is a job that runs all night at 8% utilization because the data loader is the bottleneck, and bills you in full for the privilege.
Utilization and VRAM are the two numbers that answer the questions people actually ask an hour into a run: is this machine busy, and did my batch size fit?
A missing reading says so
Not every GPU can report every field. Virtualized cards in particular often expose no temperature or power sensor at all. When that happens the metric reads Unavailable and its chart shows a gap: it does not render as 0, and it does not drag an average down toward zero.
That distinction is deliberate. A dashboard that prints 0°C for a card with no thermal sensor is worse than one that admits it has no reading, because you cannot tell a real idle GPU from a blind one. A genuine zero (a GPU truly at 0% utilization) is still shown as zero, because that is a fact worth knowing.
Shared with your team
Metrics follow the same scope as the deployment itself. Any member of the team can open the Metrics tab on any of the team's running deployments, not just the person who launched it, so "is that box anyone is still using?" is a question someone can answer without waiting for its owner to reply.
gpu utilization
GPU utilization is shown per GPU on the Metrics tab of a running deployment, as a current value and as a chart over the session. It answers the first question people ask an hour into a run: is this machine busy? A job that sits at 8% utilization because the data loader is the bottleneck still bills in full, which is why it is worth watching next to host CPU utilization.
gpu metrics
For every GPU on the box: utilization, memory used against total, temperature, power draw, and SM and memory clock speeds. The same view covers the host, with CPU utilization and system memory. A reading the card cannot report is marked Unavailable instead of shown as zero.
what is gpu load
GPU load is the utilization reading: how busy the GPU is right now. A genuine zero means the GPU is truly idle and is shown as zero. A low reading while host CPU utilization is pinned at 100% points at data loading, not at the GPU, which is why the Metrics tab shows the two together.
gpu node utilization
Utilization is reported for every GPU on a deployment, next to the host CPU and system memory, so you can read the whole node from one tab. Any member of the team can open the Metrics tab on any of the team's running deployments, so checking whether a box is still in use does not wait on its owner.
FAQ
Which GPU metrics does Aquanode show?
Per GPU: utilization, memory used and total, temperature, power draw, and SM and memory clock speeds. Alongside them, host CPU utilization and system memory. Each is shown as a current value and as a chart across the session.
How current are the numbers?
The console refreshes them every few seconds while the deployment is active, so the view tracks a running job closely rather than showing a stale snapshot.
Why does temperature or power say Unavailable?
Because that particular card cannot report it. Virtualized GPUs often expose no thermal or power sensor: a vGPU-based card reports neither. Rather than print a zero that reads as a real measurement, Aquanode marks the reading Unavailable and leaves a gap in the chart. Utilization, memory and clock speeds are read across NVIDIA supply regardless.
Can my teammates see the metrics for a box I started?
Yes. Metrics are scoped to the team that owns the deployment, so any member can open them on any of the team's running deployments.
Do I need to install anything to get metrics?
No. Metrics are collected by the Aquanode agent that runs on the box, and the Metrics option is on by default when you deploy. Leave it on and the tab is populated once the instance is running.
Can I see metrics after a deployment has stopped?
The Metrics tab is a live view of a running deployment. Once an instance stops, there is no live view to show. If you need a record beyond the session, capture what you need while the box is up.
Related reading: H100 pricing and specs, H100 vs H200, A100 vs H100, H100 and H200: SXM vs NVL vs PCIe, and The best GPUs for AI, ranked.
