At the Hot Chips conference in Cupertino ( USA) Microsoft presented a new hardware platform, which allows you to speed up the program, working on the basis of artificial intelligence. It’s called Project Brainwave and, according to Microsoft, should improve the speed of machine learning models.
The company blog It says that the platform was developed in order to ensure the operation of the AI in real time – that is, the system will respond to requests as soon as they are received, with very little delay. A faster AI is needed to process search queries, video, or, for example, interact with users, note in Microsoft.
The Project Brainwave system consists of three main layers: a high-performance distributed architecture; Hardware engine DNN, built on FGPA; Compiler and runtime library.
It is noted that Brainwave allows cloud-based deep learning models to work effectively together with the massive infrastructure on the FPGA that Microsoft has installed in its data processing centers in the last few years. This means that AI elements in applications will receive faster support from
At the conference, the accelerator was presented on Intel’s 14-nanometer FPGA Stratix 10. And at the moment the performance of the chip is estimated at 39.5 TLFOPS, although, as noted in the company, in the future it is planned to increase the performance to 90 TLFOPS (improving Stratix 10).
Now Project Brainwave supports such models of building and learning of neural networks, like Microsoft Toolkit and Google Tensorflow. In this case, both notes the Microsoft engineer Doug Burger is working on compatibility with other tools (for example, Caffe).