zentilia - Fotolia

Microsoft preps datacentres for AI revolution through deployment of reprogrammable chips

Microsoft claims field programmable gate array chips are now live in Azure datacentres in 15 countries, as it future-proofs its infrastructure for the rise of big data applications

Microsoft has deployed reprogrammable computer chips in Azure datacentres across 15 countries to ensure its infrastructure can efficiently process the data generated by its artificial intelligence (AI) activities at scale.

The software giant provided attendees at its Ignite conference in Atlanta, Georgia, with an update as to how its multi-year bid to create cloud environments with performance unburdened by the limits of Moore’s Law and the capacity of its underlying CPUs progressing.

Speaking to Computer Weekly at Ignite, Doug Burger, a distinguished engineer in Microsoft’s research division, said Moore’s Law should hold up for the next couple of years, but may struggle as the demand for big data applications grows.

“We’ve been on this 50-year exponential climb where the cost of computing has been dropping and it’s crazy, and now we’re starting to see things tap off,” he said.

“The CPUs are not getting faster at the rate they used to, it’s getting harder to scale the silicon nodes, and the demands of a lot of these big data applications are growing much faster than the underlying compute capability.”

This realisation has prompted a number of companies to focus their research and development efforts on the creation of “post-CPU” technologies as a long-term, sustainable alternative.

“With Moore’s law, each generation has a load of hard problems to solve and so they use one-time tricks to keep the scaling going, but each time it’s a different load of one-time tricks,” Burger added.

“Until now, they’ve kept the scaling going without changing the underlying paradigm, but now the one-time tricks are getting too hard and we have to think about things that change the underlying paradigm, such as programmable hardware.”

Leading the field

Microsoft’s take on reprogrammable chips are built using field programmable gate array (FPGA) technology, and – according to Burger – this means the organisation can alter the functionality of the hardware “on the fly” and in production.

The technology is central to the delivery of the organisation’s AI strategy, and should ensure its cloud infrastructure is equipped to cope with the performance demands this will place on it.

“This architecture we’ve built effectively embeds an FPGA-based AI supercomputer into our global hyperscale cloud. We get awesome speed, scale and efficiency. It will change what’s possible for AI,” said Burger during an Ignite keynote presentation.

Burger conducted some demonstrations to showcase how much faster a server featuring a single 30-watt, FPGA could carry out a photo-classification task compared with a conventional, CPU-based cloud server.

He then repeated the task by seeing how fast a FPGA-based setup could translate 1,440 pages of War and Peace from Russian to English compared with one featuring 24 CPU cores, which users a fifth more power.

The FPGA-based translation task took 2.5 seconds, while the CPU setup took eight times longer to complete the same task.

Burger said the technology has been in development at Microsoft since 2011, and has taken several generations to perfect so it could be deployed at scale in Azure and at an application level in Bing.

The Azure and Bing product groups are the only ones to have publicly announced their use of FPGAs, but – in the case of the firm’s search engine  – the technology is allowing the service to make more efficient use of its underlying resources.

Read more about computer chips

“The way we’ve architected is so the FPGA can talk to each other across the network without the CPUs being involved,” he said.

“So, if I have a server with an FPGA in it and the service running on that server is not using it, because of the architecture we defined, we can go and harvest that FGPA for other uses.”

Its use inside the Azure infrastructure means Microsoft can lay claim to being the first company to roll out FPGA at scale to create an “enhanced and intelligent cloud”.

“Certainly some companies have announced parts that are specific for one function, such as networking, but it’s really an augmentation of the server. What we have created is a flexible thing that you could use for networking, but also lots of other things,” said Burger.

“The real question to ask is when do these companies go to scale? When are they saying they’re enhancing their whole cloud and not just these three racks of servers? We’re clearly in the lead, because nobody else has announced a post-CPU cloud at scale,” he said.

Read more on Infrastructure-as-a-Service (IaaS)