Less of Moore's Law?
By Michael Fiorito, MDS
In 1965 Gordon Moore, then at Fairchild Semiconductor, claimed that number of transistors in a dense integrated circuit would double approximately every two years. Moore projected that this rate of growth would continue for at least another decade. He then revised the forecast in 1975 to doubling every two years. This observation, dubbed Moore’s Law, has been used in the semiconductor industry to guide long-term planning and to set targets for research and development.
Researchers were working to extend Moore’s Law, exploring entirely new chip materials and design techniques. But one researcher at Microsoft, Doug Burger, had another idea: Why not shuttle some of the load onto specialized chips, instead of relying on the plodding progress of the microchip industry?
Working with a few other chip researchers inside Microsoft, Mr. Burger began exploring new hardware that could accelerate the performance of Bing, the company’s internet search engine.
Mr. Burger and his team explored several options but eventually settled on chips that could be reprogrammed for new jobs on the fly, or Field Programmable Gate Arrays, (F.P.G.A.). Imagine a chip that could be programed to execute machine learning algorithms. Imagine if the same chip could then be reprogramed to handle logic that sends the many millions of data packets across its computer network.
Microsoft started to install the reprogrammable chips in 2015. Now, just about every new server loaded into a Microsoft data center includes one of these newfangled chips. Your Bing searches likely involve using this new chip technology. They are also used to speedily transport information across Azure’s massive network of underlying systems.
But there’s more.
In fall 2016, another team of Microsoft researchers built a neural network that could recognize spoken words more accurately than the average human could.
Xuedong Huang and his fellow Microsoft researchers had trained their speech-recognition service using large numbers of specialty chips - supplied by Nvidia - rather than relying heavily on ordinary Intel chips. Using special purpose chips allowed the researches to achieve the performance required to enable their solution.
Systems that rely on neural networks can learn largely on their own can evolve more quickly than traditional services. They are not as reliant on engineers writing heaps of code that explain how they should behave.While the move to reprogrammable chips is happening mostly inside massive data centers, it is only a matter of time before the innovations extend down to mobile devices, cameras, and driverless cars.
Microsoft is revolutionizing the future of chip technology. Moore’s Law may have served us well, but may not able to predict the future any more.