THE HARDWARE RACE IS ON
You can run machine learning applications on pretty much any processor. However, it thrives best on a chip that can churn out huge numbers of simple calculations with a low power consumption, a similar task to cryptocurrency mining. Traditional x86 processors aren’t particularly efficient at this. A popular choice has been Nvidia’s Tesla GPU, hardware that lends itself very nicely to the task. This has grabbed a decent slice of the emerging market for servers running neural networks. However, things are changing rapidly.
Google has decided to go all-in and develop its own hardware. In 2016, we learned of the Tensor Processing Unit, designed specifically for neural networks, and initially a fairly crude 8-bit chip. The next year, it divulged the secondgeneration version, with much improved capabilities, managing 45 Tflops. In May this year, a third-generation chip was announced, which is twice as powerful. Google is ramping up fast. Google’s TPUs are not available commercially, but the company will let you rent access via the cloud. Intel, meanwhile, has the Nervana chip due next year, the first fruits of its acquisition of Nervana Systems in 2016. This is a purpose-built machine learning chip. It has promoted the Xeon Phi, a x86 multicore server chip, as machine learning hardware, but it was always something of a stopgap. IBM has the Power9 chip, available to third parties
or via the IBM cloud services. The world’s fastest computer, Summit, contains 9,216 of them.
Elsewhere, Apple has its neural engine, part of the A11 processor for its iPhone. Amazon has a department working on AI designs, although it is unlikely to produce its own hardware. ARM is also developing new ground-up specialist neural network hardware designs. Qualcomm has its Neural Processing Engine SDK for its Snapdragon chips, too.
There are also said to be 40 or more start-up companies jumping on to this particular hardware bandwagon. Machine learning has prompted a hardware arms race that, in the long term, could be as important as the x86 race.