NVIDIA Jetson TX2 Delivers Twice the Intelligence to the Edge

Originally published at: https://developer.nvidia.com/blog/jetson-tx2-delivers-twice-intelligence-edge/

Figure 1: NVIDIA Jetson TX2 embedded system-on-module with Thermal Transfer Plate (TTP). Today at an AI meetup in San Francisco, NVIDIA launched Jetson TX2 and the JetPack 3.0 AI SDK. Jetson is the world’s leading low-power embedded platform, enabling server-class AI compute performance for edge devices everywhere. Jetson TX2 features an integrated 256-core NVIDIA Pascal…

Looks amazing Dustin!

Great article, thanks!

Wow, that is just great .... I have tried TensorRT already thanks to your other great post and it did outperform YOLO about 3 times on TegraX1. Looking forward to have TX2 n see how fast it does the same experiments 😍 So excited lol

Amazing! 4kp60 Encoding. My dreams come true!

Wow, Nice Article, Thanks!

Please correct me if my calculations are wrong: it seems like nominal ALU efficiency of GoogLeNet execution is only about 50 to 70% of the hardware peak, depending on the batch size. Is there a rogue or missing multiplication factor in the following back-of the-envelope estimation logic somewhere :

Single GoogLeNet inference == 3.16 GFLOPs of "nominal" multiplications or additions, assuming all convolutions are done by a variation of direct time-domain, approach not Winograd convolution or any other technique to reduce the amount of actual computation.

Assuming 1 "core" = FMA32x1 = FMA16x2 = 4 FP16 FLOP, peak FP16 gigaflops for Jetson TX2 at maximum clock = 256 cores * 4 FP16 FLOPs/core * 1.302 GHz = 1333 FP16 GFLOP/s (mentioned in the post as "more than a TFLOP/s of performance")

From this 1) GoogLeNet efficiency @ batch = 2: 201 FPS * 3.16 GFLOP / 1331 GFLOP/s = 48%, 2) GoogLeNet efficiency @ batch = 128: 290 FPS * 3.16 GFLOP / 1333 GFLOP = 69% 3) AlexNet efficiency @batch = 2, 250 FPS * 1.3 GFLOP / 1333 GFLOP/s = 25% 4) AlexNet efficiency @batch = 128: 692 FPS * 1.3 GFLOP / 1333 GFLOP/s = 68%


I have NVidia geforce 1080 on my windows 10 desktop.. could i run an Ubuntu VM on hyper-v that runs this software and can process stuff from a webcam or streamed from another video source like a locally networked android phone? just for learning dont have the 599 right now.

This is definitely a supercomputer among SoMs however this tiny thing is half a size and also packs a punch: