Microsoft's 'Brainwave' makes real-time AI reality

Posted Aug 24, 2017 by James Walker
Microsoft has announced a new deep learning acceleration platform that offers real-time AI. The hardware platform is a significant step forward for cloud-based neural networks. It allows deep learning networks to be streamed with "ultra-low" latency.
Microsoft Brainwave
Microsoft Brainwave
In a blog post on the technology, Microsoft explained Brainwave is made of three component systems. The foundation is a high-performance distributed system architecture for fast local computing. This is supplemented with a hardware DNN engine and a built-in compiler for trained AI models.
Brainwave's intended to make it easier for developers to run their machine learning models. They'll be able to deploy their models onto the specialised silicon. It offers greater performance than a regular processor or graphics card as it's running on a specially-designed chip that's backed by the cloud.
Brainwave's currently based on Intel's Stratix 10 field programmable gate array (FGPA) chip. Microsoft has equipped its datacentres with FGPAs over the past few years. They offer a platform for AI computing that allows developers to customise servers to the code they'll run. The FGPA is responsible for managing the AI service throughout its entire lifetime.
In the case of Brainwave, a developer could load a trained machine learning model onto the FGPA. The dedicated hardware's then able to compute insights from the AI, without requiring further training. If the model is complex and requires more power, the FGPA's scalable architecture allows it to work with another chip.
Intel Stratix 10 FGPA
Intel Stratix 10 FGPA
READ NEXT: Microsoft's AI as good as humans at voice recognition
Microsoft calls Brainwave "real-time AI" because it has hardly any latency. It's designed for AI services which need to process data as fast as it becomes available. This kind of technology has direct applications to fields such as voice recognition and autonomous vehicles. Lowering the latency to the cloud will make many forms of AI more feasible for real world use.
"Project Brainwave achieves a major leap forward in both performance and flexibility for cloud-based serving of deep learning models," said Microsoft. "We designed the system for real-time AI, which means the system processes requests as fast as it receives them, with ultra-low latency. Real-time AI is becoming increasingly important as cloud infrastructures process live data streams, whether they be search queries, videos, sensor streams, or interactions with users."
Microsoft said it intends to make Brainwave available to developers as part of its Azure cloud platform. It's also using Brainwave technology to power aspects of its consumer products. When you search with Bing or talk to Cortana, you're indirectly accessing Brainwave's real-time AI.
Microsoft said it wants Azure to offer "industry-leading" AI capabilities and "record-setting performance." Technology such as Brainwave will help Microsoft differentiate its products from rivals such as Google's TensorFlow, itself based on custom hardware.