Microsoft Gets in the AI-First Hardware Game with Project Brainwave
Microsoft has been steadily building one of the most complete artificial intelligence(AI) suites in the market. In just a few years, the Redmond giant has added relevant capabilities in areas such as machine learning platforms(Azure ML), AI APIs(Cognitive Services), R distributions(R Server), deep learning frameworks(Cognitive Toolkit), digital assistants(Cortana) and several others. A few days ago, Microsoft announced its intentions to venture into the AI-first hardware space by unveiling Project Brainwave during the Hot Chips Conference in Cupertino.
Project Brainwave is a new type of programmable chip optimized for the execution of deep learning models. The project builds on Microsoft’s previous announced field programmable gateway(FPGA) chips which offer significant advantages over traditional GPUs and CPUs when comes to the execution and scalability of deep learning programs.
FPGAs have been optimized for the execution of several types of deep learning algorithms such as convolutional neural networks(CNNs), long-short term memory(LSTM) models, reinforcement learning and a few others. The new hardware model has reported over 10x performance improvement over similar GPU and CPU architectures and has been widely deployed across Microsoft’s data science infrastructure.
Deep neural network(DNN) algorithms deployed onto FPGA chips are loaded into the hardware’s memory and scaled across different boards. Currently, Brainwave supports models built on Microsoft Cognitive Toolkit and Google’s TensorFlow but support for new deep learning frameworks should be added in the future.
Diving into chip design might seem like a deviation from Microsoft’s software core strength but it is important to realize that hardware acceleration is an increasingly relevant element of AI solutions. Additionally, Microsoft is not delivering Brainwave on its own and has reportedly partnered with Intel’s Altera on the design and manufacturing of the chips.
Some Observations About Hardware Accelerated AI
The hardware accelerated AI space is getting really crowded with software vendors such as Microsoft, Google or IBM getting into a market traditionally dominated by chip manufacturers like Intel, NVidia, ARM or Qualcomm. I wrote down a few observations that might help better understand this nascent AI market:
1 — Multi-Runtime vs. Uni-Runtime Chips: As the market evolves, we should see the adoption of both chips like FPGA that can run AI programs in multiple runtimes(Cognitive Toolkit, TensorFlow) as well as chips like Google’s TPU optimized for a single runtime (TensorFlow).
2 — Algorithm-Optimized Chips: In the future, we could also see chips optimized to execute specific deep learning algorithms such as CNNs, LSTM or reinforcement learning models.
3 — Bad News for Chip Manufacturers Stocks: In the long run, the entrance of companies like Google Microsoft, Facebook or IBM into the hardware accelerated AI market can harm the public market sentiments towards chip manufacturers. After all, Wall Street places a lot of importance in the role that IA will play in the market adoption of new technologies from the traditional chip manufacturers.
4 — Open Source Hardware Accelerated AI: We should soon see new open source designs of AI-first chips that get support from different software and hardware companies.
5 — Cloud Distributions Matter: Cloud infrastructures such as AWS, Azure, Google Cloud, Bluemix or Alibaba Cloud are likely to become the main distribution channel for AI-first chips. As a result, the cloud incumbents are in a unique position to influence the future of hardware accelerated AI.