The success of Nvidia and its new computing chip signals rapid change in IT architecture

Feb 25th 2017

“WE ALMOST went out of business several times.” Usually founders don’t talk about their company’s near-death experiences. But Jen-Hsun Huang, the boss of Nvidia, has no reason to be coy. His firm, which develops microprocessors and related software, is on a winning streak. In the past quarter its revenues increased by 55%, reaching $2.2bn, and in the past 12 months its share price has almost quadrupled.

A big part of Nvidia’s success is because demand is growing quickly for its chips, called graphics processing units (GPUs), which turn personal computers into fast gaming devices. But the GPUs also have new destinations: notably data centres where artificial-intelligence (AI) programmes gobble up the vast quantities of computing power that they generate.

Soaring sales of these chips are the clearest sign yet of a secular shift in information technology. The architecture of computing is fragmenting because of the slowing of Moore’s law, which until recently guaranteed that the power of computing would double roughly every two years, and because of the rapid rise of cloud computing and AI. The implications for the semiconductor industry and for Intel, its dominant company, are profound.

Things were straightforward when Moore’s law, named after Gordon Moore, a founder of Intel, was still in full swing. Whether in PCs or in servers (souped-up computers in data centres), one kind of microprocessor, known as a “central processing unit” (CPU), could deal with most “workloads”, as classes of computing tasks are called. Because Intel made the most powerful CPUs, it came to rule not only the market for PC processors (it has a market share of about 80%) but the one for servers, where it has an almost complete monopoly. In 2016 it had revenues of nearly $60bn.

This unipolar world is starting to crumble. Processors are no longer improving quickly enough to be able to handle, for instance, machine learning and other AI applications, which require huge amounts of data and hence consume more number-crunching power than entire data centres did just a few years ago. Intel’s customers, such as Google and Microsoft together with other operators of big data centres, are opting for more and more specialised processors from other companies and are designing their own to boot.

Nvidia’s GPUs are one example. They were created to carry out the massive, complex computations required by interactive video games. GPUs have hundreds of specialised “cores” (the “brains” of a processor), all working in parallel, whereas CPUs have only a few powerful ones that tackle computing tasks sequentially. Nvidia’s latest processors boast 3,584 cores; Intel’s server CPUs have a maximum of 28.

The company’s lucky break came in the midst of one of its near-death experiences during the 2008-09 global financial crisis. It discovered that hedge funds and research institutes were using its chips for new purposes, such as calculating complex investment and climate models. It developed a coding language, called CUDA, that helps its customers program its processors for different tasks. When cloud computing, big data and AI gathered momentum a few years ago, Nvidia’s chips were just what was needed.

Every online giant uses Nvidia GPUs to give their AI services the capability to ingest reams of data from material ranging from medical images to human speech. The firm’s revenues from selling chips to data-centre operators trebled in the past financial year, to $296m.

And GPUs are only one sort of “accelerator”, as such specialised processors are known. The range is expanding as cloud-computing firms mix and match chips to make their operations more efficient and stay ahead of the competition. “Finding the right tool for the right job”, is how Urs Hölzle, in charge of technical infrastructure at Google, describes balancing the factors of flexibility, speed and cost.

At one end of the range are ASICs, an acronym for “application-specific integrated circuits”. As the term suggests, they are hard-wired for one purpose and are the fastest on the menu as well as the most energy-efficient. Dozens of startups are developing such chips with AI algorithms already built in. Google has built an ASIC called “Tensor Processing Unit” for speech recognition.

The other extreme is field-programmable gate arrays (FPGAs). These can be programmed, meaning greater flexibility, which is why even though they are tricky to handle, Microsoft has added them to many of its servers, for instance those underlying Bing, its online-search service. “We now have more FPGAs than any other organisation in the world,” says Mark Russinovich, chief technology officer at Azure, the firm’s computing cloud.

Time to be paranoid

Instead of making ASICS or FPGAs, Intel focused in recent years on making its CPU processors ever more powerful. Nobody expects conventional processors to lose their jobs anytime soon: every server needs them and countless applications have been written to run on them. Intel’s sales from the chips are still growing. Yet the quickening rise of accelerators appears to be bad news for the company, says Alan Priestley of Gartner, an IT consultancy. The more computing happens on them, the less is done on CPUs.

One answer is to catch up by making acquisitions. In 2015 Intel bought Altera, a maker of FPGAs, for a whopping $16.7bn. In August it paid more than $400m for Nervana, a three-year-old startup that is developing specialised AI systems ranging from software to chips. The firm says it sees specialised processors as an opportunity, not a threat. New computing workloads have often started out being handled on specialised processors, explains Diane Bryant, who runs Intel’s data-centre business, only to be “pulled into the CPU” later. Encryption, for instance, used to happen on separate semiconductors, but is now a simple instruction on the Intel CPUs which run almost all computers and servers globally. Keeping new types of workload, such as AI, on accelerators would mean extra cost and complexity.

If such integration occurs, Intel has already invested to take advantage. In the summer it will start selling a new processor, code-named Knights Mill, to compete with Nvidia. Intel is also working on another chip, Knights Crest, which will come with Nervana technology. At some point, Intel is expected also to combine its CPU’s with Altera’s FPGAs.

Predictably, competitors see the future differently. Nvidia reckons it has already established its own computing platform. Many firms have written AI applications that run on its chips, and it has created the software infrastructure for other kinds of programmes, which, for instance, enable visualisations and virtual reality. One decades-old computing giant, IBM, is also trying to make Intel’s life harder. Taking a page from open-source software, the firm in 2013 “opened” its processor architecture, which is called Power, turning it into a semiconductor commons of sorts. Makers of specialised chips can more easily combine their wares with Power CPUs, and they get a say in how the platform develops.

Much will depend on how AI develops, says Matthew Eastwood of IDC, a market researcher. If it turns out not to be the revolution that many people expect, and ushers in change for just a few years, Intel’s chances are good, he says. But if AI continues to ripple through business for a decade or more, other kinds of processor will have more of a chance to establish themselves. Given how widely AI techniques can be applied, the latter seems likely. Certainly, the age of the big, hulking CPU which handles every workload, no matter how big or complex, is over. It suffered, a bit like Humpty Dumpty, a big fall. And all of Intel’s horses and all of Intel’s men cannot put it together again.