The new accelerator for machine-learning models promises to deliver performance far beyond what is possible on an unaided Pi.
Performing AI-powered tasks such as image recognition using the $35 Raspberry Pi is about to get much easier.
Google has revealed a USB stick that dramatically accelerates the rate at which the Pi and other low-power computers run trained machine-learning models.
Google's says the Edge TPU Accelerator will allow devices to run multiple state-of-the-art computer vision models on high-resolution video at more than 30 frames per second .
That level of performance would be far beyond what an unaided Pi is capable of, and seemingly above the levels of performance reported using existing AI accelerators like Intel's Movidius Neural Compute Stick.
Google knows how to design chips to handle machine learning, having had years of experience building TPU (Tensor Processing Unit) accelerators for its datacenters.
SEE: Hardware spotlight: The Raspberry Pi (Tech Pro Research)
The caveat is that the Edge TPU Accelerator would be bottlenecked by the USB 2.0 ports found on Raspberry Pi models. That slowdown would likely be exacerbated by the Pi's reliance on a shared bus for USB and Ethernet.
The stick will use the Google Edge TPU co-processor, an Application Specific Integrated Circuit (ASIC) designed to accelerate machine-learning models built using the TensorFlow Lite framework. It will connect via USB Type-C and run on Debian Linux and Android Things operating systems.
Machine-learning models will still need to be trained using powerful machines or cloud-based infrastructure, but the Edge will accelerate the rate at which these trained models can run and be used to infer information from data, for example, to spot a specific make of car in a video or to perform speech recognition.
While AI-related tasks like image recognition used to be run in the cloud, Google is pushing for machine-learning models to also be run locally on low-power devices such as the Pi.
In recent years Google has released both vision and voice-recognition kits for single-board computers under its AIY Projects program. Trained machine-learning models available to run on these kits include face/dog/cat/human detectors and a general-purpose image classifier.
Google is also releasing a standalone board that includes the Edge TPU co-processor and that bears a close resemblance to the Raspberry Pi.
The credit-card sized Edge TPU Dev Board is actually smaller than Pi, measuring 40x48mm, but like the Pi packs a 40-pin expansion header that can be used to wire it up to homemade electronics.
The Edge TPU Dev Board includes an NXP i.MX 8M Cortex-M4 processor, 1GB of RAM and 8GB of eMMC flash storage. It offers a wide variety of ports, including two USB Type-C, a USB 3.0 Type-A host, a micro-SD Card slot, a 3.5mm audio jack, two PDM MEMS microphones, a full-sized HDMI, as well as a 39-pin FFC connector for a MIPI-DSI display, and a 24-pin FFC connector for a MIPI-CSI2 camera.
Both devices are due out this fall, and those interested can sign up to be notified of their release here.