Intel’s Movidius Myriad 2 Vision Processing Unit (VPU) powers the Google Clips, a miniature camera that captures video clips and images automatically with the help of artificial intelligence (AI). The Google Clips relies on on-device hardware to capture and process the images, unlike other services of the search giant which utilizes the company’s cloud infrastructure for its machine learning computational needs. Through on-device processing, the camera can work even if there is no internet connection available, and the captured images and videos can be accessed much faster. According to the semiconductor firm, the VPU is designed to provide offer more than 1 teraflops of computational power within a power envelope of one watt. The chip’s ability to support proprietary architectures permitted Google engineers to port its algorithm into the camera. In addition, the small form factor of the chip allows it to fit in mobile devices and wearables.
The chip sports a heterogeneous architecture comprised of 12 128-bit Vector SHAVE processors and numerous hardware accelerators. The former is used for machine vision while the latter is important for processing the video and images captured by the camera. The VPU also includes 512KB of L2 cache and 2MB of on-chip memory, and it is fabricated using the 28nm low-power process node.
This is not the first time that Google developed a machine learning solution wherein the model is modified within the device. This solution is dubbed as "federated learning", and it is already used to revise the query suggestion model of the company's keyboard app, Gboard. This setup processes the model in the device and only the summary of the changes are uploaded to the Google's servers. The benefits of the modified model should be immediately experienced by the user since they no longer need to wait for Google to release an update to the model. The company unveiled the Google Clips in its fall hardware event and it includes a 12-megapixel camera equipped with a 130-degree wide angle lens. It also includes an LED flash and a physical shutter button for manually capturing the images. The device has 8GB of onboard memory and it can transfer images to the smartphone through the Wi-Fi network. There is no word yet on the tech company will start shipping the device, although Google already announced that the camera is priced at $249.