Pretty much anybody in technology knows about CPUs (central processing units), the backbone of computers. And most people also know about GPUs (graphical processing units), which excel not only at graphics but also at some complex AI algorithms. But some of us were scratching our heads when Microsoft last month touted its use of FPGAs (field-programmable gate arrays), saying the chips had already been incorporated into “the first AI supercomputer.”
Today Microsoft shed more light on FPGAs and how it plans to use them, highlighting a paper from Microsoft Research titled “A Cloud-Scale Acceleration Architecture.”
Through the work of a team dubbed Catapult, the chips are already being used to speed up web searches on Bing and to improve performance and hardware efficiency on Azure, Microsoft said in a detailed blog post today.
FPGAs can be programmed directly, rather than relying on less-efficient software to put them through their paces. And they can be quickly reprogrammed to respond to new advances or to meet unexpected needs. In contrast, it could take two years or more for chips with different specs to be designed and put into use, Microsoft said.
FPGA’s are not new. They’ve been around since the early 1980s, according to some histories.
“But until recently, no one had ever seriously tried to use them at large scale for cloud computing,” Microsoft said in the blog post. Nor had FPGAs been used as a frontline processor, according to the paper. Where once they were relegated to performing tasks sent to them, now they are the first to encounter every message going into a server.
Azure CTO Mark Russinovich said that Azure engineers can build on the work that Bing and research engineers are doing, and that in fact it was the Bing team’s success that gave him the confidence to jump on the FPGA bandwagon.
Widespread use of FPGAs could “vastly speed up advances in automatic translation, accelerate medical breakthroughs and create automated productivity tools that better anticipate our needs and solve our workday problems,” Microsoft said.