Google Unveils Microprocessor For Machine Learning Algorithms

Google Unveils Microprocessor For Machine Learning Algorithms
Google Unveils Microprocessor For Machine Learning Algorithms
Anonim
Image
Image

Google introduced the Edge TPU coprocessor, designed to hardware accelerate the work of trained neural network models. The company's engineers also developed two ready-made processor-based devices - a single-board computer, as well as a USB module for connecting to other computers.

Since the training of neural network algorithms requires large computing resources, this stage is often carried out on cloud or local servers, which allow you to quickly train the algorithm on a large amount of data. But even the application of an already trained neural network model on user devices can be a problem. On the one hand, devices are not always connected to the Internet and can transfer data processing to cloud servers, and on the other hand, many common computers do not have enough power to process data with neural network algorithms in real time. Because of this, many technology companies began to develop specialized computing devices for hardware acceleration of neural network algorithms.

Google is also developing similar devices. At the I / O conference in 2016, the company unveiled the first generation of its tensor processor (TPU), and at two subsequent conferences presented new versions of it. Google uses them only in its cloud service designed to train and execute neural network algorithms. The company has now introduced a separate Edge TPU coprocessor for hardware acceleration of already trained algorithms in end devices. It is a special purpose integrated circuit (ASIC) optimized for efficient execution of neural network algorithms.

The company does not disclose the technical details of the coprocessor, but as an example, it said that it will be able to process high-resolution video at a rate of 30 frames per second in real time, using several commonly used neural network models for this. As the main purpose of the coprocessor, Google sees smart sensors that can not only collect data for transmission to a more powerful device, but also independently carry out primary processing and make decisions.

Image
Image

Single Board Computer with Removable Module with Edge TPU

Image
Image

USB connected device with Edge TPU

Google presented not only the chip itself, but also two ready-made devices based on it. One is an expansion single board computer that contains an NXP i. MX 8M processor, Edge TPU coprocessor, and other components. It is equipped with many ports for connecting additional devices and is designed to be used as a full-fledged computer for operating various devices and prototypes. In addition, Google unveiled a device that also contains an Edge TPU, but acts as a hardware acceleration module that connects to other computers via USB. For example, it has holes on its body that are compatible with the mounting holes of the Raspberry Pi Zero single board computer.

Both devices are designed to work with the TensorFlow Lite machine learning framework developed by Google to execute neural network algorithms on mobile devices. They also support Linux and Android Things operating systems. The devices were developed as part of the AIY project aimed at amateur projects using machine learning. Earlier, the company introduced the first two devices in this project - cardboard kits for voice command recognition and image recognition using a camera.

Popular by topic