Microsoft Brainwave Accelerates Deep Learning With FPGA

Microsoft unveiled Brainwave, a FPGA -based program. It is a hardware which is capable of boosting the way machine learning models function. Microsoft is accelerating deep learning by embedding a technology based on the neural network into its programmable chips. Microsoft named this project of building artificial intelligence as Project Brainwave. It is the mixture of DPU or DNN processing units into its FPGAs. The real time artificial intelligence system developed by the Microsoft company processes the request as soon as it receives them with ultra low latency. They are hoping that by focusing on deep learning, they could adapt its infrastructure faster.  By using Intel Stratix 10 FPGAs, Brain wave can easily sustain 39.5 Teraflops.

This real time artificial Intelligence is becoming extremely important as cloud infrastructures process live data stream which includes videos, queries, sensor, streams and interactions with users. Microsoft has been investing in FPFA for quite a long time now. All the algorithms are written into FPGAs which makes them really efficient. Brainwave uses an integrated circuit called Field programmable gate arrays in Azure to do many operations. Microsoft says that by using FPGAs it is making changes in its approach to run predictions. 

Also Read: DeepMind Hassabis Now Supports Graphcore AI Chips

Earlier it was all about creating generalized CPUs, now it is all about creating custom chips that excel at the particular task. As a result of which, the chips have been made for machine learning. It is believed that the major effort has been Google’s Tensor Processing Unit. A lot of tools are required to do deep learning via Brainwave. Both, Microsoft TensorFlow and Cognitive Toolkit are supported by Microsoft Brainwave. Brainwave converts existing models without even knowing how much pressure it creates when porting existing models. Microsoft ensures the speed improvement with the launch of Brainwave as the FPGAs is attached directly to the network in the data centre.