Renowned US electronics manufacturer and distributor Adafruit announced that it is working on its own USB Tensor Processing Unit (TPU) powered by a TPS62827 Coral accelerator chip. This chip could increase the processing power of your next Raspberry Pi-powered Machine Learning project.
Adafruit's Coral TPU board will come as a USB device. When inserted into a Raspberry Pi 4, Adafruit says that Coral will improve the performance of TensorFlow Lite inferences by up to ten times in certain circumstances. It should offer similar performance to Google's Coral USB accelerator, which uses the same chip, and costs $59, but is not in a stick form factor.
We recently reviewed Adafruit's BrainCraft HAT, designed as a single module to aid Raspberry Pi 4-powered Machine Learning projects. While the BrainCraft HAT had no TPU and was unable to help process the data, it provided the means to operate a device that could collect and output data.
Processing was handled via the Raspberry Pi 4's CPU, which manages around 4-5fps. With this new Adafruit TPU, we can expect the fps to jump to over 20fps!
We asked Limor Fried, CEO Adafruit for more information.
"The goal is for this new chip to be used with Pi 4 computers specifically to greatly speed up vision recognition projects - we found that Mobilenet SSDlite type demos ran at maybe 1~2 fps and with this add on we could get to a real-time friendly 20fps which would allow many more edge AI projects to use the common Pi 4."
There's no firm word yet on pricing or availability as Fried is still working on the price of materials needed to manufacture the stick. However, she said that she's hoping to offer it for somewhere around the price of a Raspberry Pi 4 (though she didn't say which Raspberry Pi 4) and would have more news in a few weeks.
It is still early days, so we have no details on price, but it is hoped that Adafruit's Coral will retail for less than Google's Coral TPU, possibly around the same price as a Raspberry Pi 4.