The Coral Accelerator Module enables big AI in a small package

Smaller than a US penny, the solderable multi-chip module brings the power of Google's Edge TPU to your fingertips.

ic_share_24px Created with Sketch.
Close Created with Sketch.

From the beginning, Coral has offered development boards and self-contained AI accelerators to developers and prototypers. Now, Google’s platform for on-device AI brings a new form factor to the table, one that enables a wide range of production applications.

Coral engineers have packed the Google Edge TPU machine learning co-processor into a solderable multi-chip module (MCM) that’s smaller than a US penny. The new Accelerator Module lets developers solder privacy-preserving, low-power, and high performance edge ML acceleration into just about any hardware project.

Flexible and affordable

The Accelerator Module complements Coral’s lineup of USB and PCIe Accelerators without the encumbrance or footprint associated with USB cables and PCIe connectors.

The module comprises a tiny circuit board with an RF-shielding metal lid and contains all of the power and interface circuitry needed to run the Edge TPU and features USB 2.0 and PCIe interfaces.

The Accelerator Module is:

  • Small enough, at 15.0 x 10.0 x 1.5 mm, to fit into almost any design.

  • Lets developers solder the Edge TPU acceleration onto a PCB.

  • Saves even more space, thanks to a lack of connectors or cables.

  • Requires no exotic boards or surface mounts, thanks to conventional rather than high-density interconnector parts.

  • Provides a standard USB 2.0 or PCIe interface to an application processor.

  • Reduces costs with Coral’s lowest price point.

  • Allows developers to add a single system component to their designs to enable ML acceleration.

“We put the power management IC into the MCM,” explains Coral senior hardware engineer James McLurkin. “You add three external capacitors, then connect USB or PCIe, and you’re done. You can get back to the rest of your design by lunchtime.”

The Accenture and Google Partnership

The Accenture team works closely with its customers to source the right sensors, processors, and other components for its AI solutions.

For situations where bandwidth is constrained, privacy is paramount, or processing speed is of the essence, Accenture brings Coral to the table — a natural extension of Accenture’s ongoing work with Google.

The Accenture Google Business Group is a partnership between Accenture and Google focused about 80% on the Google Cloud Platform, according to Ghanathe. “I lead the industry innovation practice. What that means is we take Google’s technology and create industry solutions.”

For local AI, the Accenture team always recommends Coral to clients. “The major advantage with Google’s Edge TPU is that it’s cost-effective,” Sudhindra says. “The second one is they’re open-source, so they have a robust developer community.”

Notes from the field

Developers have already been testing pre-release Accelerator Modules. They are finding that it fills a critical need for them and their customers.

For Siana Systems, this means not only prototyping for development purposes but also creating production-ready designs to aid in relaying the value AI-enabled solutions have to offer. “You need something that looks like a finished product,” says Sylvain Bernard, Siana owner, founder, solution architect, speaking of the prototypes he presents to customers in the retail, industrial, and health care sectors. “The Accelerator Module fits perfectly into what we are doing.”

Bernard also appreciates the flexibility the Coral platform brings to the table with TensorFlow Lite, including the ability to easily develop solutions on a PC before porting them to stand-alone devices. Most of his company’s designs are for vision processing systems, including the MPCam, which has the Coral Accelerator on board.

Modular hardware developer Gumstix has begun incorporating Coral components into boards for robotics, vision, and IoT applications.

“Recently, we have seen many low power devices offering on-device (edge) AI capabilities,” says Andrew Smith, Gumstix manager, Product Development and Engineering, Smart Manufacturing. “While cool and exciting, many of these new low power devices are not suited for high data rate applications. Adding a Coral component could potentially transform a traditional processor into a well-equipped edge-AI device that can function in demanding applications, such as those with high data rates.”

The small form factor of the Accelerator Module lets Coral-powered devices go even more places, Smith says. “We like giving designers the opportunity to minimize board size whenever possible.” Towards that, Gumstix has opened preorders of its new Raspberry Pi CM4 UprevAI board, featuring the Coral Accelerator Module.

The Mustang-T100-T5 from IEI will feature five Coral Accelerator Modules on a single PCIe card

Edge AI hardware provider IEI also values the ability to provide low power, low cost solutions for sophisticated AI applications. The high performance and small size of the Accelerator Module enable the Mustang-T100-T5, coming in early 2021.

“Our design is a high-density AI accelerator product using five cores on a single board” says product manager Brian Chen. The card will fit five Accelerator Modules into a single PCIe slot. “We have a lot of demand to implement AI” Chen explains. Specifically demand for edge AI is growing and it’s important to have a product that works well with prior investments in legacy systems. “So, a product that offers a lot of AI acceleration in a PCIe add on card is suitable.”

The Accelerator Module’s integrated power control circuits reduce costs while providing powerful AI capabilities to a wide range of smart city and factory automation applications.

The SDT-A9X4 Quad-Core SoM with the Coral Accelerator Module is the recommended choice by SDT for embedded AI-enabled machine vision devices

This development board features the SDT-A53X8 octa-core SoM with the Coral Accelerator Module is the recommended for embedded machine vision devices with rich UI display requirements

IoT platform developer SDT Inc. sees edge AI’s resource-saving capabilities as a game-changer for its system-on-modules used for sound, vision, and other pattern recognition applications.

“We no longer have to drain our CPU resources for solving ML problems. The fact that the Coral platform works seamlessly with TensorFlow Lite means that Google not only provides the chip, but it also one of the best ML development environments and resources,” explains Jiwon Yune, CEO of SDT.

“We think that the computing power of each Coral module is great for lots of popular AI problems. In the past, end-users wanted something simple, like the fastest car, the smallest semiconductor, or cameras with the highest resolution. Now customers are seeking energy efficient, eco-friendly and intelligent solutions to accompany those products. They know that embedding AI on devices can prepare them for the next generation.”

Yune likes to install as many AI processors as needed for a given application. Parallel inference and model pipelining capabilities on the Coral platform combined with the small form factor of the Accelerator Module brings that to the table. “If a problem is too complicated, just adding one or two additional modules can do the job,” Yune says.

Looking ahead

Small size. Low cost. Easy development on multiple platforms. Flexible capabilities that developers can add or subtract on their PCBs as needed, including modules running in parallel to scale ML capabilities for any use case. These features of the Coral Accelerator Module promise to bring edge AI to the next level.

“We’re enabling a diversity of hardware form factors with this module," says Billy Rutledge, director of Coral. "Acceleration is the key component for edge AI. And this addition to our product lineup allows our ecosystem partners to build products better suited to their customers' needs.”