Skip to content

Latest commit

 

History

History
64 lines (37 loc) · 2.61 KB

eTPU.md

File metadata and controls

64 lines (37 loc) · 2.61 KB
Raw

external VPU and TPU

In the Robolab two types of external VPU/TPUs are available, which can be used for mobile machine learning applications. This are the Intel Myriad X VPU and the Google Coral Edge TPU.

In addition, an external GPU ( eGPU.md) and four workstations with good internal GPUs are available as rapid prototyping machines (README.md).

This external VPU and TPU are meant to be able to use pretrained models on your mobile device (iOS or Android) or your mobile robot (Rasberry Pi or Linux), because coprocessors are typically not available on those mobile devices. The downside is that you have to do cross-development on your laptop and port your code than to the mobile device.

VPU Hardware Specifications

  • Type: Intel® Movidius™ Myriad™ X Vision Processing Unit (VPU)
  • Cpu: none
  • Memory: none
  • VPU for calculation: Myriad™ X MA2480 - 16 SHAVE cores
  • interface: USB 3.0 Type-A socket

More details: https://software.intel.com/content/www/us/en/develop/hardware/neural-compute-stick.html

VPU Interface

https://docs.openvinotoolkit.org/latest/openvino_docs_install_guides_installing_openvino_raspbian.html

VPU Tutorials

https://software.intel.com/content/www/us/en/develop/articles/get-started-with-neural-compute-stick.html

TPU Hardware Specifications

  • Type: Coral - Google Edge TPU
  • Cpu: none
  • Memory: none
  • TPU for calculation: Global Unichip Corp
  • interface: USB 3.1 (gen 1) Type-C socket

More details: https://coral.withgoogle.com/products/accelerator/

TPU Interface

https://coral.withgoogle.com/tutorials/accelerator/#setup-for-linux-or-raspberry-pi

https://www.tensorflow.org/lite

https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite

TPU Tutorials

https://www.tensorflow.org/guide/using_tpu

https://coral.withgoogle.com/tutorials/edgetpu-retrain-classification/

https://coral.withgoogle.com/tutorials/edgetpu-retrain-classification-ondevice/

TPU Performance

Note the performance in Frequent Asked Questions, where the external TPU boosted the performance of a desktop CPU with a factor 20x.

Notes

When you first set up the Coral USB Accelerator, you can select whether to use the default or maximum clock frequency. The maximum clock frequency runs at 2x the default setting.

To change the setting later, you currently must uninstall the libedgetpu_*.so file, then rerun the install script, in which you'll be prompted to select the setting.