2411839_en-US

[#1]

How to get NPU Temperature and utilization for NXP i.MX93 EVK

I am using i.MX93 EVK and want to monitor the Ethos-U NPU temperature & utilization(load) at runtime how can I get this data ?

The agenda is to monitor the HW-parameters during the inference of AI-models.
I could monitor the below parameters using i.MX8MPlus I want to achieve the similar for i.MX93.

• 8MPlus
  • RAM=$(free -m | awk '/Mem:/ {print $3}')
  • CPU_LOAD=$(vmstat 1 2 | tail -1 | awk '{print 100 - $15}')
  • CPU_TEMP=$(awk '{print $1/1000}' /sys/class/thermal/thermal_zone0/temp 2>/dev/null || echo "N/A")
  • GPU_TEMP=$(awk '{print $1/1000}' /sys/class/thermal/thermal_zone1/temp 2>/dev/null || echo "${CPU_TEMP}")
  • GPU_LOAD=$(cat /sys/kernel/debug/gc/load 2>/dev/null | awk '/load/ {gsub("%","",$3); if (max=="" || $3>max) max=$3} END {if (max!="") print max "%"; else print "N/A"}')

• I.mx93
  • RAM=$(free -m | awk '/Mem:/ {print $3}')
  • CPU_LOAD=$(vmstat 1 2 | tail -1 | awk '{print 100 - $15}')
  • CPU_TEMP=$(awk '{print $1/1000}' /sys/class/thermal/thermal_zone0/temp)

How can I monitor the GPU_LOAD & GPU_TEMP for i.MX93 npu(ethos-u) ?

LinuxRe: How to get NPU Temperature and utilization for NXP i.MX93 EVK

Hi @sonal-acldigital 

Neither the i.MX 8M Plus nor the i.MX 93 offers a method to monitor NPU temperature in real time.

- imx8MP, thermal_zone0 represents the cpu-thermal temperature ，thermal_zone1 represents the soc-thermal temperature.

- imx93, thermal_zone0 represents the cpu-thermal temperature.

- NPU, There is no such tool like nputop.

For information on performance tuning, please refer to Section 7.1.2, "Profiling for Hardware Accelerators," in the ML User Guide.

B.R

Re: How to get NPU Temperature and utilization for NXP i.MX93 EVK

@pengyong_zhang Thanks for clarifying the thermal zones for the i.MX 8M Plus and i.MX 93. Could you please clarify the following:

i.MX8MPLUS
1. GPU Temperature: Is there a dedicated system path (like /sys/class/thermal/thermal_zoneX/temp) to read individual GPU temperatures, or is it covered solely under the soc-thermal zone (thermal_zone1)?
2. GPU Load: Is reading /sys/kernel/debug/gc/load the standard method for checking GPU load, or is the process detailed in Section 7.1.2 of the ML User Guide required for this as well?

i.MX93:
Since there is no native tool (like nputop) to monitor NPU load and temperature, we would like to ask: Is it valid to derive relative NPU utilization by measuring the reduction in CPU load when offloading inference execution to the NPU vs. running on the CPU alone?
If this method isn't recommended, is there any alternative driver node or performance counter we can use to estimate NPU activity during an inference run?