The Detected Computing Power Is Temporarily Lower Than the Target Value During the Computing Power Test Performed by the ascend-dmi Tool

Possible Causes

When the device fails to dissipate heat in a timely manner, the current and power consumption increase, causing temperature increase. As a result, EDP throttling occurs and the detected computing power is lower than the target value of the Ascend AI Processor. This is a protection mechanism of Ascend AI Processors.

This problem may occur temporarily when the device starts to run and the temperature increases instantaneously. Atlas 300T training cards (model 9000) deliver high computing power and high power consumption. Therefore, throttling is more likely to happen during the computing power test.

Troubleshooting

Run the npu-smi info -t common -i id command to query the processor frequency. If the current frequency Aicore curFreq(MHZ) is less than the nominal frequency Aicore Freq(MHZ), the problem is caused by EDP throttling.