Google has announced that it is to launch a hardware product based on its Tensor Processing Unit (TPU) deep-learning acceleration technology, aimed at bringing artificial intelligence smarts to the Internet of Things (IoT): Edge TPU.
While Google announced its first custom-built Tensor Processing Unit (TPU) hardware back in May 2016, and has been releasing news about upgraded versions every year since, the company has thus far made the acceleration hardware available only internally, first to power its own services and then as an option through its cloud computing platform. In a surprise announcement made late last night, though, the company confirmed that it was ending that exclusivity by launching its first commercial TPU-based product: The Edge TPU.
'Edge TPU is Google’s purpose-built ASIC [application-specific integrated circuit] chip designed to run TensorFlow Lite ML models at the edge,' explains Injong Rhee, vice president for Internet of Things (IoT) and Google Cloud. 'When designing Edge TPU, we were hyperfocused on optimising for "performance per watt" and "performance per dollar" within a small footprint. Edge TPUs are designed to complement our Cloud TPU offering, so you can accelerate ML training in the cloud, then have lightning-fast ML inference at the edge. Your sensors become more than data collectors—they make local, real-time, intelligent decisions.'
The Edge TPU, considerably less powerful than its server-centric brethren, will initially launch as part of a Raspberry Pi-like system-on-module (SOM) development kit based on an NXP Semiconductors system-on-chip processor. The bare TPU chip itself, meanwhile, is due to appear in products from companies including NXP, Arm, Accton, Nokia, Smart Catch, and Trax, among others named as partners in the programme.
At the same time, Google has announced Cloud IoT Edge, an expansion to Google Cloud which allows machine learning models built on the full-fat cloud-based TPUs to be executed on the Edge TPU along with GPU- and CPU-based accelerators running Android Things or any other Linux-based operating system.