Different Results on Different TX2s

I’ve deployed three Jetson units, two devkits and one Connectech Rudi TX2. They are all using the same neural network confirmed by checking the hashes of the files. When I run the same image set on all three units I get different results from each unit. If I run the image set again on the same TX2 it will reliably produce the same results.

Is this normal? Sometimes the detections vary by more than 100%. For example, Connecteh Rudi will detect 1258 objects. With the same images and the same network on one of the devkits I get 692 objects detected.

There are minor software version differences. Can the version differences below really be causing this?

ConnectTech Rudi
nvidia@tegra-ubuntu:~$ apt list --installed | grep cuda

WARNING: apt does not have a stable CLI interface. Use with caution in scripts.

cuda-command-line-tools-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-core-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cublas-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cublas-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cudart-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cudart-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cufft-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cufft-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-curand-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-curand-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusolver-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusolver-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusparse-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusparse-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-documentation-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-driver-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-libraries-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-license-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-misc-headers-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-npp-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-npp-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvgraph-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvgraph-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvml-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvrtc-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvrtc-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-repo-l4t-9-0-local/now 9.0.252-1 arm64 [installed,local]
cuda-samples-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-toolkit-9-0/unknown,now 9.0.252-1 arm64 [installed]
libcudnn7/now 7.0.5.13-1+cuda9.0 arm64 [installed,upgradable to: 7.0.5.15-1+cud 9.0]
libcudnn7-dev/now 7.0.5.13-1+cuda9.0 arm64 [installed,upgradable to: 7.0.5.15-1 cuda9.0]
libcudnn7-doc/now 7.0.5.13-1+cuda9.0 arm64 [installed,local]
libgie-dev/unknown,now 4.0.4-1+cuda9.0 arm64 [installed]
libnvinfer-dev/unknown,now 4.0.4-1+cuda9.0 arm64 [installed,automatic]
libnvinfer-samples/unknown,now 4.0.4-1+cuda9.0 arm64 [installed,automatic]
libnvinfer4/unknown,now 4.0.4-1+cuda9.0 arm64 [installed,automatic]
nv-tensorrt-repo-ubuntu1604-ga-cuda9.0-trt3.0.4-20180208/now 1-1 arm64 [installed,local]
tensorrt/unknown,now 3.0.4-1+cuda9.0 arm64 [installed]
nvidia@tegra-ubuntu:~$

DevKit 1
nvidia@iceman:~$ apt list --installed | grep cuda

WARNING: apt does not have a stable CLI interface. Use with caution in scripts.

cuda-command-line-tools-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-core-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cublas-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cublas-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cudart-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cudart-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cufft-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cufft-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-curand-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-curand-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusolver-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusolver-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusparse-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-cusparse-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-documentation-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-driver-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-libraries-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-license-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-misc-headers-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-npp-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-npp-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvgraph-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvgraph-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvml-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvrtc-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-nvrtc-dev-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-repo-l4t-9-0-local/now 9.0.252-1 arm64 [installed,local]
cuda-samples-9-0/unknown,now 9.0.252-1 arm64 [installed,automatic]
cuda-toolkit-9-0/unknown,now 9.0.252-1 arm64 [installed]
libcudnn7/now 7.1.5.14-1+cuda9.0 arm64 [installed,local]
libcudnn7-dev/now 7.1.5.14-1+cuda9.0 arm64 [installed,local]
libcudnn7-doc/now 7.1.5.14-1+cuda9.0 arm64 [installed,local]
libgie-dev/now 4.1.3-1+cuda9.0 arm64 [installed,local]
libnvinfer-dev/now 4.1.3-1+cuda9.0 arm64 [installed,local]
libnvinfer-samples/now 4.1.3-1+cuda9.0 arm64 [installed,local]
libnvinfer4/now 4.1.3-1+cuda9.0 arm64 [installed,local]
tensorrt/now 4.0.2.0-1+cuda9.0 arm64 [installed,local]
nvidia@iceman:~$

Hi,

How do you install your device?
Are they using the same JetPack installer?

Also, could you maximize the device performance to see if helps?

sudo ./jetson_clocks.sh

Thanks.