Identify the source first
A CPU's instantaneous clock, effective clock and SMU table clock come from different measurement paths. Effective clock also reflects activity during a sampling interval. A BIOS voltage setting, VID and measured rail voltage must be kept separate too.
For memory, I distinguish AOD firmware settings from the output voltage measured by a PMIC. An unavailable sensor should not become a zero: that looks like a successful measurement. Its availability needs to be explicit in the interface.
The PM table version matters
An SMU PM table is decoded through a field map. The CPU family and PM version determine how that map should be interpreted. Copying an index from one version to another can produce a plausible number with the wrong meaning.
The size of a raw memory read is also different from the size of the table itself. A larger read does not turn bytes beyond the table boundary into sensor fields. Checking bounds and units matters as much as checking the resulting values.
What the local comparison proves
In the local review on 1 October 2026, I matched ZenStation readings on a Ryzen 5 3400G with nearby HWiNFO records. Refresh intervals and background load differed, so small differences were not automatically errors. Voltage, temperature and power were compared with their own sources and units.
Total GPU memory usage and dedicated memory usage are different fields. Adding a multiplier to force agreement would hide that distinction. Similarly, a motherboard input named AUXTIN by a provider is not automatically a physical CPU temperature sensor.
This physical comparison covers the 3400G. Source reviews and saved reports exist for other Ryzen families; they should not be described as physical validation on every device.