Skip to content
TELEMETRY / NOTE 007

Chip telemetry as an input for cooling control.

Servers already report temperatures, power draw and utilization for their own protection. This note looks at what that telemetry contains, how it can inform facility cooling and what has to be true before a controller relies on it.

arflo Engineering · · 4 MIN

What the hardware reports.

Modern accelerators and servers expose their state through standard interfaces. On NVIDIA GPUs, the management library and its command-line tool report values such as GPU temperature, power draw, utilization and fan speed. Baseboard management controllers offer server-level sensors, including inlet temperature and power, through interfaces such as Redfish.

Which values are available, at what resolution and how often they update depends on the device, its firmware and its configuration. An inventory of the actual signals is the first step of any integration.

What the values mean.

A reported GPU temperature comes from sensors at specific locations. It is not a full map of the die, and the hottest point can be warmer than the reported value. Vendors define their limits in terms of their own sensors, which makes the reported value suitable for tracking margin to those limits.

Power draw is the more direct signal for cooling. Almost all of it becomes heat, and it changes before any temperature does. Utilization is a coarser indicator: the same utilization can correspond to different power levels depending on the workload.

Three uses in cooling control.

As a feedforward signal, power draw tells the cooling system about a load change when it happens. As a constraint, the margin between component temperatures and their limits shows how much room there is to raise supply temperature or lower flow. As validation data, measured temperatures show whether a thermal model predicts the real system correctly.

Facility sensors remain necessary. They describe the conditions the hardware cannot see: supply and return temperatures, flow, pressure and the state of the cooling plant.

What has to be true first.

Timing: values from different sources need timestamps that can be aligned, and the controller has to know how old each value is. A stale reading that looks current is worse than a missing one.

Access: IT telemetry and facility controls usually belong to different teams and networks. Reading the data requires an agreed, secured path, and in colocation environments the tenant, not the operator, owns the server data.

Fallback: a controller that uses telemetry needs defined behavior for the case that it disappears. Losing the signal should return the system to a conservative operating mode, not leave it at an aggressive setpoint.

From signal to model.

Telemetry alone is a set of readings. Its value for control comes from the relationship between them: how power, coolant state and temperatures depend on each other over time. arflo builds that relationship as a thermal model from chip to infrastructure and adapts it with measured data. How well the model fits is established by comparing its predictions with these same measurements on the specific system.