FØCAL is a profiling and automated test farm platform based on Docker and LTTng for testing Linux edge AI software on the BeagleBone, Raspberry Pi, Jetson, Up Squared, and Google Coral.

A venture-backed startup called FØCAL has launched a cloud-based test farm of the same name designed for hardware/software codesign of Linux-based edge AI and computer vision applications. The test farm offers testing on common Linux hacker boards for a flat price of $0.10 an hour

FØCAL allows hardware vendors to “expose their products to key software workflows before the hardware goes to fabrication,” says the company. The FØCAL tooling “exposes finicky bare metal to software workflows that were previously only available on the cloud.”

Jetson Nano Dev Kit

Current devices on the testing farm include the BeagleBone Black, Google’s i.MX8M and Edge TPU based Coral Dev Board, the Raspberry Pi 3B+, and Aaeon’s Intel Apollo Lake based Up Squared SBCs. The farm also provides Nvidia’s CUDA-enhanced Jetson modules and dev kits, including the Jetson TX2, TX1, Xavier, and Nano.

The boards are deployed in managed device clusters at secure data centers around the U.S. It appears that you can order the test boards to be customized with AI add-ons such as the UP Squared board’s optional UP AI Core X modules. Supported Linux distros include Ubuntu, Debian, Arch, NVidia L4T, Google Mendel, Raspbian, and more.

Coral Dev Board

Major components of FØCAL include:

• f0cal/farm — This “AWS for devices” provides an API client and command line interface wrapper that allows developers to “spin up devices” on the FØCAL device farm “as if they were VMs.” The device farm integrates support for the CircleCI cloud-native continuous integration and delivery platform.
• f0cal/my-device — This CLI front-end to various vendor-supplied OS installers and manual install procedures provides Docker-like workflows for bare metal devices.
• f0cal/my-code — A user-friendly frontend to the open source LTTng application tracing framework connects user-space code to kernel events, hardware counters, and external sensors for software performance visualization. The frontend “makes it easier to create, share, and leverage tracepoint definitions for third-party libraries and facilitate test-time assertions against hardware, software, and kernel conditions.”

As explained in several Medium blog essays by founder Brian Rossa, a computational neuroscientist formerly at Lockheed Martin – Advanced Technology lab, edge AI and computer vision development is plagued by a chicken-or-egg problem between hardware and software that begs for a codesign solution. Whereas with cloud-based AI, the hardware is generally a very well-known factor, edge AI algorithms run differently on various hardware platforms, each with different and usually cutting edge (i.e. unknown) hardware accelerators.

“It’s extremely difficult to know in advance what the actual latency and power consumption benefits of any given hardware accelerator will be, especially if algorithm selection isn’t finalized,” writes Rossa. “Without a detailed knowledge of both the algorithms and the larger application architecture, hardware acceleration is opaque and undifferentiated. No one person on the team has enough context to evaluate the hundreds of seemingly identical options.”

Automated testing platforms such as FØCAL can help solve potential problems with a given platform, says the company. In the process, FØCAL can also help development teams determine which combination of CPUs, GPUs, NPUs, and other accelerators are the best match for a given software application — and why. FØCAL also enables and encourages hardware and software developers to work together on the common problem rather than argue and procrastinate in their separate silos, says FØCAL.

 
Further information

The FØCAL platform is available for free download on GitHub. The FØCAL testing service is available at $0.10 an hour. More information may be found at the FØCAL website.