Performance evaluation and energy consumption of neural networks on Nvidia Jetson TX2.
Nvidia Jetson TX2 has CPU ARMv8: 2x Denver2 + 4x A57 + 1x GPU Pascal 256 cores + 8 GB DDR4 RAM.
Consumption reading is done through INA3221 shunt type sensors integrated on the board.
The different operating modes are analyzed using the nvpmodel command that uses Tegrastats.
Mode | Name | ARMv8 Denver (Cores) | Freq GHz | ARMv8 A57 (Cores) | Freq. Ghz | Pascal GPU GHz |
---|---|---|---|---|---|---|
0 | MAX N | 2 | 2.0 | 4 | 2.0 | 1.30 |
1 | MAX Q | 0 | 4 | 1.2 | 0.85 | |
2 | MAX P Core All | 2 | 1.4 | 4 | 1.4 | 1.12 |
3 | MAX P ARM | 0 | 4 | 2.0 | 1.12 | |
4 | MAX P Denver | 1 | 2.0 | 1 | 2.0 | 1.12 |
Nvidia TensorRT environment with PMLib for energy measurement. Then, the analyzed Convolutional Neural Network (CNN) have been AlexNet. GoogleNet-12, Inception-V3 and VGG-19 with ImageNet dataset.