Multiple Edge TPU Demo

This demo shows how to use model pipelining (with multiple Edge TPUs) to process a video with a larger model.

This demo allows two configurations, one for the Dev Board and one for generic platforms (including x86_64, ARM64, and ARMv7). The Dev Board configuration is optimized to run efficiently for the platform. The generic version relies on SW rendering and decoding to ensure compatability.

Requirements

Dev Board Requirements

• Dev Board
• 1 or 2 USB Accelerators
• USB Accelerators should be connected to Dev Board with a powered USB hub to guarantee sufficient power supply

For the Dev Board, This demo is tested on Mendel Eagle with Eel2 release

Generic Platforms

• A x86_64, ARM64, ARMv7 host with multiple EdgeTPUs attached (via USB, PCIe, or both).

For generic platforms, this is tested on a Linux x86_64 host with USB accelerators.

Compiling the demo

You will need to install docker on your computer first to compile the project.

In order to build all CPU targets type:

make DOCKER_TARGETS=demo docker-build

You can also specify the CPU targets you'd like:

make DOCKER_TARGETS=demo DOCKER_CPUS="k8 aarch64" docker-build

The binary should be in out/$(ARCH)/demo directory.

Running the demo

To run the demo, copy test_data and generated binary multiple_edgetpu_demo to your device. Run:

./multiple_edgetpu_demo --num_segments 3 \
  --video_path <your video>

You can enter ./multiple_edgetpu_demo --help to see all the flag options and default values.

Performance expectation

To analyze a 720p 60 FPS video with inception_v3 model, the system can process about 20 FPS with 1 Edge TPU; and 50+ FPS with 3 Edge TPUs using model pipelining. Performance may vary in the generic configuration based on SW decoding and rendering ability.

License

Apache License 2.0