Google LLC’s Coral product line is a family of tiny chip modules that hardware makers can integrate into their devices to provide machine learning features. Today, ahead of next week’s Consumer Electronics Show in Las Vegas, the search giant added several new products to the series.

All the offerings are based on the Edge TPU. It’s a miniaturized version of the Tensor Processor Unit, an application-specific integrated circuit optimized for AI that Google uses in its data centers. The Edge TPU trades away most of the original TPU’s horsepower for efficiency gains and is small enough to fit on a penny but still packs a surprisingly big punch.

The first new Coral product Google debuted today is the Accelerator Module. It delivers an Edge TPU inside a fairly bare-bones hardware case that can be installed on a gadget’s circuit board in four different ways, including via USB and PCIe. It aims to be a no-frills option for hardware makers looking to bake Google silicon into their products. 

Google also unveiled two new versions of its existing Coral System-on-Module product, which represents a step up on the functionality ladder. It also packs a single Edge TPU but in a much bigger module that comes with a central processing unit, graphics chip, memory and a preinstalled Linux distribution. The fact that it’s a self-contained solution reduces development times for hardware makers, providing the ability to bring AI devices to market faster. 

The new versions of the Coral System-on-Module feature 2 and 4 gigabytes of LPDDR4 memory, respectively. The original model has 16 gigabytes.

Lastly, Google is rolling out the Dev Board Mini, which is likewise a smaller version of an existing Coral product. It’s not meant to be incorporated into commercial products but is rather intended as a platform for hardware teams to prototype machine learning features. The Dev Board Mini has a single Edge TPU, a CPU, a GPU, 2 gigabytes of RAM, 8 gigabytes of flash storage and Linux built-in.

The new Coral modules’ onboard Edge TPU can handle up to 4 trillion operations per second. According to Google, that’s enough to run multiple computer vision models at 30 frames per second, or a single model at up to 400 frames per second. 

“More and more industries are beginning to recognize the value of local AI, where the speed of local inference allows considerable savings on bandwidth and cloud compute costs, and keeping data local preserves user privacy,” Coral Director of Research Billy Rutledge wrote in a blog post. He added that companies have in the past year harnessed the modules for applications “across a broad set of industries that range from healthcare to agriculture to smart cities.”

Photo: Google