JetPack 2.3 with TensorRT Doubles Jetson TX1 Deep Learning Inference

Originally published at:

Deep Neural Networks (DNNs) are a powerful approach to implementing robust computer vision and artificial intelligence applications. NVIDIA Jetpack 2.3, released today, increases run-time performance of DNNs in embedded applications more than two-fold using NVIDIA TensorRT (formerly called GPU Inference Engine or GIE). With up to 20x higher power efficiency than an Intel i7 CPU during inference workloads, NVIDIA’s 1…


What is the procedure to benchmark Caffe GoogleNet/AlexNet based on TensorRT on Jetson TX1 using Jetpack 2.3?

Hello, the Caffe procedure is available from this sticky post and TensorRT includes timings example in the samples installed on the Jetson with JetPack 2.3.

Thanks for your reply. The only samples I found were for classification and I could run those. But how do I get forward time or images/sec?

Check the timing sample located at /usr/src/gie_samples/samples/sampleGoogleNet , it installed along with TensorRT / JetPack 2.3.


I still can not reproduce the benchmark of TensorRT 21fps detection as paper. I got only 8 fps on this detection

./detectnet-camera ped-100


Hi VuNguyen, the 21 figure is the number of GoogleNet (ImageNet) images per second per watt. The raw number of GoogleNet images per second is 203fps. What you are referring to is DetectNet, which requires additional layers and computation. GoogleNet is image recognition and DetectNet is multi-class multi-object localization. Since the publication of this article, there has been additional work and progress on faster DetectNet however for a future release.

Hi Dustin Franklin,
Good to hear that. Do you know the plan for a future release?

Can artificial neural netowrk problems be solved using above kit?