Jetson_benchmark Minimum memory requirements

lrnxd9120 · October 11, 2023, 9:02am

Hi All,
The minimum memory requirement for Jetson Benchmark is how much?
I can get result from Orin NX 8G but I got 0fps and “Error opening engine file” in Orin Nano 4G. I also saw the low memory warning.

Is the memory status normal?

Thanks.

AastaLLL · October 12, 2023, 2:44am

Hi,

Which command are you using for running on the Orin Nano?

Based on your log, there is still some memory available in your environment.
Could you share the complete log with us?

Thanks.

lrnxd9120 · October 12, 2023, 5:42am

My step about execute jetson_benchmark:
$ sudo apt install nvidia-jetpack
$ sudo unzip jetson_benchmarks.zip
$ cd jetson_benchmarks
$ sudo sh install_requirements.sh
$ sudo mkdir models
$ sudo python3 utils/download_models.py --all --csv_file_path benchmark_csv/orin-nano-benchmarks.csv --save_dir models
$ sudo python3 benchmark.py --all --csv_file_path benchmark_csv/orin-nano-benchmarks.csv --model_dir $PWD/models/ --jetson_clocks

And i got result:
Model Name FPS
0 inception_v4 0.000000
1 vgg19_N2 0.000000
2 super_resolution_bsd500 0.000000
3 unet-segmentation 0.000000
4 pose_estimation 0.000000
5 yolov3-tiny-416 0.000000
6 ResNet50_224x224 0.000000
7 mobilenet_v1_ssd 1154.754585
8 ssd_resnet34_1200x1200 0.000000

inception_v4 log:
inception_v4_b32_ws2048_gpu.txt (4.9 KB)
vgg19_N2 log:
vgg19_N2_b64_ws2048_gpu.txt (4.9 KB)
super_resolution_bsd500 log:
super_resolution_bsd500_b8_ws2048_gpu.txt (5.1 KB)
unet-segmentation log:
unet-segmentation_b32_ws2048_gpu.txt (5.0 KB)
pose_estimation log:
pose_estimation_b64_ws2048_gpu.txt (4.9 KB)
yolov3-tiny-416 log:
yolov3-tiny-416_b32_ws2048_gpu.txt (5.0 KB)
ResNet50_224x224 log:
ResNet50_224x224_b32_ws2048_gpu.txt (4.9 KB)
ssd_resnet34_1200x1200 log:
ssd_resnet34_1200x1200_b1_ws2048_gpu.txt (5.0 KB)
dmesg:
dmesg.txt (136.0 KB)

Thanks.

AastaLLL · October 16, 2023, 5:44am

Hi,

Thanks for your testing.

We are checking with our internal team.
Will update more info with you later.

Thanks.

AastaLLL · October 23, 2023, 1:50am

Hi,

Thanks for your patience. We expect the jetson_benchmarks to work on Orin Nano.

From your log files, it seems like the models are saved in a directory named models but the path specified in the command is looking for models in a directory named modules.

Could you please try again after changing the last 2 commands to:

sudo python3 utils/download_models.py --all --csv_file_path $PWD/benchmark_csv/orin-nano-benchmarks.csv --save_dir $PWD/models/
sudo python3 benchmark.py --all --csv_file_path $PWD/benchmark_csv/orin-nano-benchmarks.csv --model_dir $PWD/models/ --jetson_clocks

Please ensure that $PWD/models/ correctly points to where the downloaded models are located.

Thanks.

lrnxd9120 · October 23, 2023, 5:27am

Hi,

Sorry for confusing. The log files is the second time test,and the second time test use “modules” to save benchmark models. In fact, the command i used matches the save models path.

test with:

sudo python3 utils/download_models.py --all --csv_file_path $PWD/benchmark_csv/orin-nano-benchmarks.csv --save_dir $PWD/models/
sudo python3 benchmark.py --all --csv_file_path $PWD/benchmark_csv/orin-nano-benchmarks.csv --model_dir $PWD/models/ --jetson_clocks
Result:
0 inception_v4 0.000000
1 vgg19_N2 0.000000
2 super_resolution_bsd500 0.000000
3 unet-segmentation 0.000000
4 pose_estimation 0.000000
5 yolov3-tiny-416 0.000000
6 ResNet50_224x224 0.000000
7 mobilenet_v1_ssd 1161.069871
8 ssd_resnet34_1200x1200 17.577550
log:
vgg19_N2_b64_ws2048_gpu.txt (4.9 KB)
super_resolution_bsd500_b8_ws2048_gpu.txt (5.1 KB)
unet-segmentation_b32_ws2048_gpu.txt (5.0 KB)
pose_estimation_b64_ws2048_gpu.txt (4.9 KB)
yolov3-tiny-416_b32_ws2048_gpu.txt (5.0 KB)
ResNet50_224x224_b32_ws2048_gpu.txt (4.9 KB)

I also tried using GitHub - JetsonHacksNano/installSwapfile: Install a swap file on the NVIDIA Jetson Nano Developer Kit. This should help with memory pressure issues.. It seems slightly improved.
Result:
0 inception_v4 183.265680
1 vgg19_N2 0.000000
2 super_resolution_bsd500 0.000000
3 unet-segmentation 0.000000
4 pose_estimation 0.000000
5 yolov3-tiny-416 389.784717
6 ResNet50_224x224 634.665598
7 mobilenet_v1_ssd 1156.875674
8 ssd_resnet34_1200x1200 17.787978
log:
vgg19_N2_b64_ws2048_gpu.txt (4.9 KB)
super_resolution_bsd500_b8_ws2048_gpu.txt (5.1 KB)
unet-segmentation_b32_ws2048_gpu.txt (5.0 KB)
pose_estimation_b64_ws2048_gpu.txt (4.9 KB)

Thanks.

AastaLLL · October 24, 2023, 6:28am

Hi,

Thanks for the feedback.
We need to discuss this with our internal team.

Will get back to you later.

lrnxd9120 · October 31, 2023, 5:29am

Hi @AastaLLL ,

Are there any updates about the issue?

AastaLLL · October 31, 2023, 8:40am

Hi,

Thanks for your patience.

Our internal team is still checking on this.
Will get back to you once we get a response.

Thanks.

AastaLLL · November 1, 2023, 1:43am

Hi,

Please give the attached CSV file a try.
orin-nano-4gb-benchmarks.csv (1.3 KB)

It mainly modifies the batch size for the non-working model.
So you can also modify it based on your environment status.

Thanks.

lrnxd9120 · November 3, 2023, 7:08am

Hi,

Thank you for your suggestion.
I have tried the attchment.

below is the score :
Model Name FPS
0 inception_v4 113.736908
1 vgg19_N2 95.358279
2 super_resolution_bsd500 101.997273
3 unet-segmentation 69.097159
4 pose_estimation 221.999593
5 yolov3-tiny-416 322.847119
6 ResNet50_224x224 327.531229
7 mobilenet_v1_ssd 1048.020436
8 ssd_resnet34_1200x1200 17.603146

That can solve my issue but I observed that the result is much lower than the standard score. Is there are any suggestions that can improve it?

Thanks.

AastaLLL · November 6, 2023, 5:32am

Hi,

Did you run the test in the below environment?

https://developer.nvidia.com/embedded/jetson-benchmarks

These Benchmarks were run using Jetpack 5.1.1
Each Jetson module was run with maximum performance (Max Frequencies in MAXN for JAO64, JAO32, ONX16, ONX8; and 15W mode for JON8, and 10W mode for JON4)

Thanks.

lrnxd9120 · November 6, 2023, 6:02am

Hi,

Yes, my Jetpack version is 5.1.1 and run with 10W(Orin Nano 4GB).

AastaLLL · November 8, 2023, 7:10am

Hi,

Could you reboot the system and try the benchmark again?
Please don’t enable any disk-based swap memory since the access time is relatively longer.

Thanks.

lrnxd9120 · November 9, 2023, 6:50am

Hi,

I tried remove the additional swap size,and try the benchmark again with orin-nano-4gb-benchmarks.csv .
below is the score :
Model Name FPS
0 inception_v4 0.000000
1 vgg19_N2 0.000000
2 super_resolution_bsd500 102.007965
3 unet-segmentation 68.727215
4 pose_estimation 220.685477
5 yolov3-tiny-416 323.757111
6 ResNet50_224x224 0.000000
7 mobilenet_v1_ssd 1054.882615
8 ssd_resnet34_1200x1200 17.517354

Some process have been killed,cause by not enough memory.
dmesg as below:
benchmarkDmesg.txt (136.2 KB)

Thanks.

AastaLLL · November 9, 2023, 6:53am

Hi,

Thanks for testing.

We need to check with our internal team.
Will get back to you soon.

AastaLLL · November 13, 2023, 5:03am

Hi,

Please adjust the batch size to get the best performance.
For example, the best batch size Inception_v4 is 32.

Thanks.

lrnxd9120 · November 13, 2023, 6:25am

Hi,

Is it normal to get 0 fps for Orin Nano 4GB?
I know 32 batch size can get the best performance but now even set batch size to 1, it still got 0 fps after I remove extra swapsize.
Can you help me to check what is the expected normal performance for Orin Nano 4GB?

Thanks.

AastaLLL · November 14, 2023, 3:59am

Hi,

Please ignore the disable swap suggestion.
Just try it as the way you tried with the attached csv.

Below is the batch size and fps we got with Orin Nano 4GB.

Thanks.

system · December 5, 2023, 1:57am

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.