The Hardware page combines the current host snapshot with stored history for the devices and metrics that SoAI can collect. It is the place to check CPU, memory, storage, network, and GPU behavior while a model or plugin is running.
This page is the manual view. A model can read a live host snapshot through the hardware_snapshot
agent tool, then use SoAI's two internal hardware tools to act on the result. See
Monitor from an agent below and
Hardware tools.
Open the Hardware page
- In the SoAI WebUI, open Hardware from the side navigation.
- Review the live cards for the host's CPU, memory, storage, network interfaces, and detected GPUs.
- Use Info in the page header when you need the host information view.
-
Use Logs to continue investigating a problem in the logs page. That page requires
Log access (
LOG_ACCESS).
The page requires Hardware read access (HARDWARE_READ). A Standard user does not receive that permission by default. The process card appears only with
Process viewing (HW_PROCESS_VIEW), and GPU controls have their own GPU tuning (HW_GPU_TUNING) requirement.
Missing data means that SoAI did not receive that metric for the current device and host state (it is not a
zero measurement).
A mixed-vendor host can show NVIDIA, AMD,
and Intel cards together. SoAI probes every detected vendor and keeps device
identities distinct across vendors. On Linux, amd-smi and
xpu-smi provide management telemetry where the installed driver supports them. Native
radeon and i915 interfaces can also provide temperature and clock readings without those
tools. On Windows, supported
AMD sensors use the driver's shared-memory logging interface, while
Intel management telemetry uses xpu-smi.exe.
NVIDIA telemetry uses the vendor driver and
NVML or nvidia-smi.
A sensor can report zero; an unavailable reading means there is no current measurement. Stale readings show their age. Dedicated GPU memory and integrated shared memory are identified separately. Actual clock frequency, requested frequency, and configured limits describe different quantities.
Use charts, sampling, and comparisons
- In Chart, choose a Device.
- Choose a Metric that the selected device reports.
- Choose the time range and chart type you need.
- For OHLC and Heikin Ashi charts, use Interval to choose how much time each candle represents.
The available devices, metrics, time ranges, and intervals come from the current hardware snapshot and history capability data. If history is disabled, Chart shows History is disabled by the backend. Select Export in the page header after selecting the hardware source you want to export. SoAI opens an export preview for that source.
While hardware detection is busy, a history chart can show the last known device information with its age; plotted values still come from stored history. If the device is unavailable, wait for detection to finish and retry the series.
Drag or swipe a history chart to move through time. Use the mouse wheel or a pinch gesture to change the visible span; when you move into older history, SoAI loads more stored samples until it reaches the oldest available reading. Changing chart type keeps the visible time window. Chart detail adjusts as you zoom in or out. Periods when SoAI was not running appear as gaps labelled No data · SoAI offline.
Change sampling and retention
The Advanced control in Settings appears only to an Administrator with
Configuration changes (CONFIG_PATCH). Select the Settings button at the right of the header, then select
Settings from its menu. Select Advanced in the page header, choose the
System tab, and open the Hardware group. Change
Monitoring interval ms, Hardware history enabled, or
Db retention hours, then select Save. If Settings shows the
Restart SoAI action, use it when you are ready to apply the pending restart.
The factory defaults sample every 3000 milliseconds, enable history, and retain hardware history for 168 hours.
Their current values, defaults, and restart requirements are listed in
SYSTEM.HARDWARE.MONITORING_INTERVAL_MS,
SYSTEM.HARDWARE.HARDWARE_HISTORY_ENABLED, and
SYSTEM.HARDWARE.DB_RETENTION_HOURS.
Compare a change
Keep the model, plugin, prompt shape, concurrency, selected device, and time window consistent while comparing a change. Read temperature, power, utilization, memory use, and plugin behavior together. One metric on its own does not establish the cause of a slowdown or failure.
Monitor from an agent
hardware_snapshot returns the same live readings this page shows. Left unset it reports uptime, OS,
CPU, memory, and GPU; its components argument adds storage, network, swap, motherboard, and measured
disk and network speed. include_processes adds the per-GPU process list and requires an administrator,
the same distinction the process card makes here. The tool reads the live snapshot only. Stored history, the
Chart time ranges, and Export stay on this page.
A WebUI conversation needs a model that supports tool calling, MCP usage (MCP_USE), and hardware_snapshot selected under Chat configuration > Tools.
The tool requires an authenticated user context but does not check the Hardware page's
HARDWARE_READ permission.
include_processes=true also requires an Administrator.
Monitoring that should act on what it reads pairs it with hardware_control and
hardware_benchmark: read a baseline, change one GPU setting, then stress-test the result. Those two are
administrator-only, Execute mode only, and internal to SoAI. They belong to the administrator
catalog, so MCP server mode never publishes them to third-party software that uses SoAI as an MCP server, which can
read a snapshot but cannot change or stress a GPU.
GPU controls and tuning covers the manual equivalent of that loop, and
Hardware tools covers all three tools.