Microsoft to deploy its own reprogrammable chips for deep learning

In the coming weeks, Microsoft will be using self-designed field programmable gate arrays for deep learning applications. Ultimately, Project Catapult’s chips should lead to an artificial intelligence supercomputer.

Microsoft already uses its own FPGAs for the Bing search engine and the Azure cloud storage service, but will expand this to new search algorithms based on deep neural networks in the coming weeks, Wired writes. The reprogrammable chips should provide a significant speed improvement when running machine learning applications.

The Catapult v2 servers should be used for image recognition and language processing and in the future also for compression and encryption at Office 365. In the coming years, Microsoft will equip all its new servers with an FPGA. The new architecture must be able to be deployed more flexibly than the first generation of Catapult servers. The accelerators of those servers consist of an Altera Stratix V D5 fpga with a pci-e x16 connection, 4GB ddr3 and two qsfp interconnects.

Microsoft has been working on Project Catapult since 2011 to integrate FPGAs into its hardware platform. The company now invests 5 to 6 billion dollars annually in hardware. Microsoft’s move to reprogrammable chips was so important to Intel that it was the main reason for the chip giant to acquire Altera, Intel vice president Diane Bryant said. Intel bought that FPGA maker last year for $16.7 billion. Project Catapult will remain the driving force behind the development of a “global supercomputer” until 2030, according to Peter Lee, vice president of Microsoft Research; then the step to a quantum computer could be made.