Jetpack7.2 T5000 性能测试问题

Module: T5000

System Version: R39.2.0

Using a self-developed carrier board based on T5000. Previously, the system was adapted based on Jetpack 7.1.0 (R38.2.0), and all tests were normal. Performance tests in MAXN mode could reach 120W, which is within the normal range.

Recently, we carried out system adaptation and upgrade development for Jetpack 7.2.0 (R39.2.0). During performance testing, we found that the actual power in MAXN mode is only 50W.

We have cross-validated with an overload board, and the phenomenon is consistent: the Jetpack 7.2.0 system only reaches 50W, while the Jetpack 7.1.0 system reaches 120W.

To rule out BSP adaptation issues, we flashed an unmodified NVIDIA Jetson system onto our device and found that the phenomenon remains consistent.


Executed before performance testing

sudo nvpmodel -m 0
sudo jetson_clocks

Meanwhile, we have tested other modes besides MAXN on Jetpack 7.2.0, including 120W, 90W, and 70W, and the results were the same. The actual power reaches a maximum of only 50W; by comparing it with the digital display of the actual adjustable power supply, the data from jtop is correct.
There are also tools available for replacing performance tests, refer tohttps://forums.developer.nvidia.com/t/t5000-sample-to-use-full-performance/369329/4,The result remains unchanged。
Tested on the NVIDIA Thor kit with Jetpack 7.2.0, the performance tests are normal; the issue only occurs on our carrier board.State that the software tools for performance testing are OK

We would like to know what changes were made to the software code in Jetpack 7.2.0 compared to Jetpack 7.1.0, which led to the observed changes in compatibility.

How can we modify the BSP to adapt it to our carrier board? Thank you!

By using
sudo tegrastats

100%@649,100%@650,100%@650,100%@649] EMC_FREQ 18%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@75C/75.281C tj@75.843C/75.843C soc012@75.281C/75.281C gpu@75.843C/75.843C soc345@73.593C/73.875C VDD_GPU 4336mW/4338mW/4338mW VDD_CPU_SOC_MSS 17353mW/11690mW/18931mW VIN_SYS_5V0 11074mW/8300mW/12781mW
07-31-2026 10:58:42 RAM 27093/125810MB (lfb 5154x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@649,100%@649,100%@650,100%@650,100%@650,100%@651,100%@649,100%@649,100%@650,100%@649,100%@650,100%@650,100%@650,100%@650] EMC_FREQ 10%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@74.468C/75.281C tj@75.875C/75.875C soc012@75.031C/75.281C gpu@75.875C/75.875C soc345@73.468C/73.875C VDD_GPU 4338mW/4338mW/4338mW VDD_CPU_SOC_MSS 15776mW/11868mW/18931mW VIN_SYS_5V0 9144mW/8337mW/12781mW
07-31-2026 10:58:43 RAM 29609/125810MB (lfb 4663x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@649,100%@650,100%@649,100%@649,100%@649,100%@649,100%@649,100%@650,100%@649,100%@650,100%@650,100%@649,100%@649,100%@649] EMC_FREQ 5%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@74.531C/75.281C tj@75.781C/75.875C soc012@75.062C/75.281C gpu@75.781C/75.875C soc345@73.562C/73.875C VDD_GPU 4338mW/4338mW/4338mW VDD_CPU_SOC_MSS 15381mW/12014mW/18931mW VIN_SYS_5V0 8839mW/8358mW/12781mW
07-31-2026 10:58:44 RAM 33151/125810MB (lfb 3923x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@650,100%@649,100%@650,100%@649,100%@650,94%@650,100%@650,100%@649,100%@649,100%@650,100%@650,100%@649,100%@650,100%@650] EMC_FREQ 6%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@74.875C/75.281C tj@76.718C/76.718C soc012@74.781C/75.281C gpu@76.718C/76.718C soc345@73.968C/73.968C VDD_GPU 12615mW/4669mW/12615mW VDD_CPU_SOC_MSS 16163mW/12180mW/18931mW VIN_SYS_5V0 10464mW/8442mW/12781mW
07-31-2026 10:58:45 RAM 35886/125810MB (lfb 3268x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@649,100%@649,100%@649,100%@650,100%@650,100%@649,100%@650,100%@650,100%@649,100%@650,100%@649,100%@650,100%@649,100%@649] EMC_FREQ 9%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@75.062C/75.281C tj@77.062C/77.062C soc012@75.093C/75.281C gpu@77.062C/77.062C soc345@74.187C/74.187C VDD_GPU 14586mW/5050mW/14586mW VDD_CPU_SOC_MSS 16558mW/12349mW/18931mW VIN_SYS_5V0 10769mW/8531mW/12781mW
07-31-2026 10:58:47 RAM 38579/125810MB (lfb 2595x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@650,100%@649,100%@649,100%@649,100%@650,100%@650,100%@650,100%@649,100%@650,100%@649,100%@650,100%@650,100%@650,100%@649] EMC_FREQ 10%@4266 GR3D_FREQ @[1575,1574,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@75.125C/75.281C tj@77.093C/77.093C soc012@75.125C/75.281C gpu@77.093C/77.093C soc345@74.187C/74.187C VDD_GPU 14586mW/5403mW/14586mW VDD_CPU_SOC_MSS 16558mW/12505mW/18931mW VIN_SYS_5V0 10769mW/8614mW/12781mW
07-31-2026 10:58:48 RAM 41353/125810MB (lfb 1903x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@649,100%@649,100%@650,100%@650,100%@650,100%@649,100%@650,100%@650,100%@650,100%@649,100%@649,100%@649,100%@649,100%@650] EMC_FREQ 10%@4266 GR3D_FREQ @[1575,1575,1575] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@75.218C/75.281C tj@77.187C/77.187C soc012@75.375C/75.375C gpu@77.187C/77.187C soc345@74.468C/74.468C VDD_GPU 14981mW/5745mW/14981mW VDD_CPU_SOC_MSS 16558mW/12649mW/18931mW VIN_SYS_5V0 10769mW/8691mW/12781mW
07-31-2026 10:58:49 RAM 44194/125810MB (lfb 1193x4MB) SWAP 0/2048MB (cached 0MB) CPU [100%@650,100%@650,100%@649,100%@650,100%@650,100%@649,100%@649,100%@649,100%@649,100%@649,100%@650,100%@650,100%@649,100%@650] EMC_FREQ 10%@4266 GR3D_FREQ @[1575,1575,1574] NVENC0_FREQ @1692 NVENC1_FREQ @1692 NVDEC0_FREQ @1692 NVDEC1_FREQ @1692 NVJPG0_FREQ @1692 VIC off OFA_FREQ @1692 PVA0_FREQ off APE 300 cpu@75.25C/75.281C tj@77.281C/77.281C soc012@75.281C/75.375C gpu@77.281C/77.281C soc345@74.406C/74.468C VDD_GPU 14981mW/6064mW/14981mW VDD_CPU_SOC_MSS 16558mW/12784mW/18931mW VIN_SYS_5V0 10668mW/8759mW/12781mWnvidia@tegra-ubuntu:~$ sudo nvpmodel -q
NV Power Mode: MAXN
0

I noticed that jtop displays the CPU frequency as 2.6GHz, but in MAXN mode, the CPU frequency is 650MHz.

nvidia@tegra-ubuntu:~$ cat /sys/devices/system/cpu/cpufreq/policy0/scaling_max_freq
2601000
nvidia@tegra-ubuntu:~$ cat /sys/devices/system/cpu/cpufreq/policy0/scaling_cur_freq
2601000
nvidia@tegra-ubuntu:~$ ls /sys/devices/system/cpu/cpufreq/
policy0  policy10  policy12  policy2  policy4  policy6  policy8  schedutil
nvidia@tegra-ubuntu:~$ for i in /sys/devices/system/cpu/cpufreq/policy*; do
echo $i
cat $i/scaling_cur_freq
cat $i/scaling_max_freq
done
/sys/devices/system/cpu/cpufreq/policy0
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy10
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy12
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy2
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy4
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy6
2601000
2601000
/sys/devices/system/cpu/cpufreq/policy8
2601000
2601000
nvidia@tegra-ubuntu:~$nvidia@tegra-ubuntu:~$ for i in /sys/devices/system/cpu/cpufreq/policy*; do
echo $i
cat $i/scaling_governor
done
/sys/devices/system/cpu/cpufreq/policy0
schedutil
/sys/devices/system/cpu/cpufreq/policy10
schedutil
/sys/devices/system/cpu/cpufreq/policy12
schedutil
/sys/devices/system/cpu/cpufreq/policy2
schedutil
/sys/devices/system/cpu/cpufreq/policy4
schedutil
/sys/devices/system/cpu/cpufreq/policy6
schedutil
/sys/devices/system/cpu/cpufreq/policy8
schedutil
nvidia@tegra-ubuntu:~$

We flashed the JP7.2 system adapted for our device onto the Thor kit for verification, and the test results were also OK.

I would like to know what changes occurred from JP7.1 to JP7.2 that caused this phenomenon.
It feels as though the jp7.2.0 system has updated some detection or protection features, and because certain proprietary devices on the carrier board were not recognized, it has limited the CPU and GPU performance.

My results suggest that JetPack 7.2 through updates to Jetson Linux 39.2, the NVIDIA driver stack, and NVML, may have resolved some monitoring limitations present in earlier releases, allowing tegrastats, nvidia-ml-py/pynvml, and Jetson Power GUI to report a more complete set of statistics. And jtop continues to be updated.


Depending upon how you installed jtop

# Copy attached power.py to either:
$HOME/.local/share/jtop/lib/python3.12/site-packages/jtop/core/
find /usr/lib /usr/local/lib -iname jtop

# Then
sudo systemctl daemon-reload
sudo systemctl restart jtop
sudo jtop


# quick test, if your jtop is in /usr then  basename $HOME/.local/share/jtop/bin/python

$HOME/.local/share/jtop/bin/python - <<'PY'
from pprint import pprint
from jtop import jtop

with jtop() as jetson:
    if jetson.ok():
        pprint(jetson.power)
PY

If you install jtop with one of:

Option 4: Run jtop with or without sudo, and install without --break-system-packages
This installation method works on all Jetson Developer Kits. NOTE: For Jetson AGX Thor Dev Kit follow this
sudo -v
curl -LsSf https://raw.githubusercontent.com/rbonghi/jetson_stats/master/scripts/install_jtop_torun_without_sudo.sh | bash
Upgrade jtop
sudo -v
curl -LsSf https://raw.githubusercontent.com/rbonghi/jetson_stats/master/scripts/upgrade-jtop.sh | bash

Then once this pr makes it upstream merged into master, then you can upgrade or install jtop with:

uvx jtop-installer


power.py.txt (16.1 KB)

Thank you for your reply, but this is unrelated to the jtop code.
We copied the pinmux file from jp7.1.0 to jp7.2.0, and after reflashing the QSPI, the tests on jp7.2.0 are now normal.
It should be caused by a change in the configuration of one or more pins; we are still investigating exactly which one.
There are many changes in the content of the pinmux file from jp7.1.0 to jp7.2.0. I am not sure why this was done?
Logically, the pinmux file should remain unchanged or be slightly modified; it should not change with the version.

Please check what is the cpu frequency when no loading applied and with jetson_clocks.

You should not use MAXN to run any stress test. Most likely you hit OC3 event and leads to OC throttling.

The issue has been confirmed.
Caused by the SYSTEM_OC_N pin

@@ -50,7 +51,7 @@ pinmux@ac281000 {
            batt_oc_paa3 {
                nvidia,pins = "batt_oc_paa3";
                nvidia,function = "soc_therm_oc2";
-               nvidia,pull = <TEGRA_PIN_PULL_NONE>;
+               nvidia,pull = <TEGRA_PIN_PULL_UP>;

pin is set to UP in jp7.1, but changed to NONE in jp7.2. There is an external 1.8V pull-up on the Thor kit, but this pin is floating on our board, so an internal pull-up is required.
Mechanism: OC_N is active low. Floating + no internal pull-up → pin floats, causing the SoC to misread an “active overcurrent/overtemperature alarm,” which triggers a proactive power cap to 50W. After changing to PULL_UP, the pin is stably pulled high (= no alarm), the restriction is lifted, and power recovers to 120W. This perfectly matches the observed phenomenon.