(Image: Pierluigi Longo)
Dizzily increasing PC power used to be a given. No longer – speeds stalled a decade ago and only a radical reboot of computing will accelerate things
TEN years ago, computers stopped getting faster. Stroking your sleek smartphone or latest laptop, this may seem a rather implausible statement. Surely there’s no contest between it and a decade-old desktop?
Advertisement
That’s true – in a way. But even if computer chips weren’t made of silicon, the comparison would be built on sand. Continually increasing computing power used to rest on a solid foundation of ever smaller, faster chips. In the past decade, though, it has become a case of using more chips, less efficiently. Chip speed stalled sometime around 2004.
You don’t need to be the type who camps outside stores for the latest gizmo to be concerned. Since the silicon chip’s invention some 40 years ago, exponentially increasing computing power has become a bedrock of economic and social development. Believe the prophets and today we stand on the brink of a further information revolution, one where the ability to crunch huge data sets quickly and cheaply will provide us with a cornucopia of benefits. These include everything from better weather forecasts to personalised drugs to an “internet of things”, in which items as diverse as refrigerators and clothes will exchange streams of life-improving data in real time.
Only if the hardware can keep up. “People think ‘big data’ is going to kick in and be the second IT revolution, but you need the raw power, the chips,” says Glenn Martyna, a researcher at IBM. First, someone needs to satisfy our need for speed.
It’s 50 years ago this April that Gordon Moore, co-founder of the chip company Intel, first would roughly double every two years. More precisely, Moore said that the number of transistors in a single circuit or processor would double every 18 to 24 months.
Perhaps physicists’ greatest invention, the transistor is a switch made of a semiconductor such as silicon. These materials are normally insulators, but start conducting if you apply a small electrical voltage. This switching behaviour allows digital information to be stored and processed in the form of strings of 1s (current on) and 0s (current off).
Although chip evolution has come in fits and starts, what became known as Moore’s law proved pretty accurate (see “When the chips are down”). Intel’s “4004” silicon chip, which kicked off the PC revolution in 1971, packed in 2300 transistors, each 10 micrometres across. Today’s cutting-edge chips contain up to 5 billion, each roughly 20 nanometres across – the width of a couple of hundred atoms. There’s still some room at the bottom, too: it’s only at about a nanometre across that this sort of transistor architecture would be expected to become unreliable, thanks to quantum-mechanical effects such as electrons tunnelling through seemingly impenetrable barriers.
For most of the past half-century, the momentum of miniaturisation has delivered the goods. Smaller transistors have meant shorter distances for electrons to travel, less time to switch state and more bang for your computing buck. By the early 1980s, the first PCs were being marketed trumpeting maximum transistor-switching speeds – known as chip clock rate – measured in megahertz, or a million cycles a second. In 2000, the first gigahertz processors were . By 2003, the Intel Pentium 4 HT chip was working with a clock rate of 3 gigahertz.
Heat death
And since then? Very little. As of 2015, the fastest central processor chips available for personal electronic devices still have clock rates largely hovering around 3 GHz.
The reason is more steam age than silicon age: heat. Billions of transistors switching billions of times a second create billions and billions of tiny puffs of heat, adding up to a huge amount of waste energy that must be transported away from a tiny area. That can only happen so fast, and a decade ago we hit thermodynamic limits on the speed of heat disposal. Individual transistors have continued to get smaller, and tweaks to their architectures – for example “multi-gate” transistors that can be more sensitively controlled using more than one voltage – have made them faster and more efficient. But to avoid the chips melting or exploding, each one has had to work less, throttling the overall chip speed. “These guys are still shrinking those transistors, but the belief that clock speeds go hand in hand with this miniaturisation is now over,” says .
So it’s not that Moore’s law is broken, just that we aren’t reaping the rewards as we used to. In the past decade, firms such as Intel have kept things going by using more transistors, but engineering them so that at any one time most of them aren’t operating. “Multi-core” chips split computing tasks over different areas to provide more processor wow, while spreading the pesky power issues out.
But this approach, too, is reaching the end of the line. The more cores a processor has, the more time and energy it costs to communicate between them, eventually negating any gains.
Meanwhile, demand continues to rise apace. The IT research company Gartner has estimated that more than 25 billion smart devices will be – massive central processor farms – by 2020. Beyond that, the vision is of a nexus of devices in our hands and in central server farms that work in tandem to perform increasingly data-heavy real-time analysis of everything from our environment, health status and online behaviour to how traffic is moving and the contents of our fridges.
This amounts to a huge additional burden on chip performance. Smartphones are already getting uncomfortably warm, and server farms across the world consume as much power as Japan, says Martyna. They are increasingly moving to colder climes, taking advantage of lower ambient temperatures to dissipate chip heat more quickly. “The industry always saw this point coming, but it took a while for people to wake up to how serious it was,” says Martyna.
In the US, one reaction has been the creation of the , or STARnet. A collaboration between academic labs and major commercial chipmakers coordinated by the US government agency DARPA, it is funded to the tune of almost $200 million. It aims to look at computing in the round, finding out where efficiency can be improved all the way from underlying physical processes and materials to the smooth interaction of computer hardware and software.
Easy wins
Austin heads up one of six sub-projects, the . It is looking at ways to deal with huge volumes of data without abandoning 50 years of silicon development. One focus is on extending the reach of graphics processing units, or GPUs, which were invented to satisfy the information-intensive needs of computer game graphics. Unlike a computer’s central processor, which has to perform many varied tasks in unpredictable sequences according to a user’s demands, a GPU specialises in doing the same job – rendering the millions of pixels of a moving image – in parallel over and over again. This ability can also be used to speed up other computational problems that have a similar parallel nature, and Austin and his colleagues are looking at ways to combine multiple GPUs to make them useful for a wide range of general computing tasks.
Another potential easy win, especially in multi-core processors, is to increase transport speeds between chip components. One way to do that is to change the material used to make the connectors. In the late 1990s, the chip industry moved from aluminium to copper connectors, which were faster and much cheaper, without consumers noticing, says . Everyone’s favourite wonder material graphene – a one-atom thick form of carbon that conducts blisteringly fast – is a strong contender for the next leap. Or we could skip wires altogether and get chips to communicate using light. IBM researchers have suggested this could be a practicable way of communicating between cores by the end of the decade. Others think this is optimistic.
CFAR’s approach to the same problem is somewhat different. Many common computational tasks, for example accessing databases, require the central processor and the computer’s long-term memory to interact constantly, bouncing signals back and forth. By analysing how a computer’s software manages this data exchange, chips can be designed to minimise the distance signals travel, and thus the heat they expend – or software can be designed to make more efficient use of the chip architecture.
Such gains will be welcome, but they skirt the central issue: transistor performance. For a more radical speed-up, we might need to rethink this fundamental unit of computing.
Martyna and IBM are placing some of their money on a very different transistor technology, the basics of which have been around for a century or more: piezoelectrics. These are materials that change from conductors to insulators when subjected to mechanical strain. In 2008, Martyna was one of the first to discover that the voltage to switch the state of these materials is much lower than to switch a semiconductor such as silicon. The win is potentially a 100-fold reduction in power consumption and a 10-fold increase in processor speed.
The challenge is getting the things to work reliably at the nanoscale, where small imperfections on a material’s surface can make big differences to its performance. IBM and others are now collaborating with a series of European metrology institutions to develop the necessary technology within a €4 million EU-funded project called . It’s so far, so good, says project member of the UK’s National Physical Laboratory: switching has been demonstrated successfully at a scale of 100 nanometres, with further size reductions planned.
It’s not the only approach to reduce the power consumption of individual transistors. Another is spintronics, which takes advantage of a property of electrons called spin. Each electron has two possible spin states – up and down – which can be used to encode each bit of data. Flipping between these states takes much less power than switching a current on and off. It’s a principle that has been used since the mid-1990s to improve the performance of memory storage on computer hard drives, and another STARnet project, the , Interfaces and Novel Architectures, aims to build on prototype spin transistors to extend that success to processors. Part of the problem is to find the material that enables the most efficient spin switching, and get that switching to work at room temperature.
Another alternative is to harness the power of quantum tunnelling. This process is generally seen as a hurdle to future miniaturisation: if transistor gates get too thin, electrons can disappear from one side and reappear on the other, ruining the switching. But the ability to control the process would make it possible to switch using far less power.
Tunnelling transistors have been made in the lab, but a new generation of incredibly thin “two-dimensional” materials, among them graphene, promise much greater control. Electrons have to burrow through much less material in such structures, potentially improving switching speed. Niftier transistors have also been fashioned by introducing carbon nanotubes into conventional transistors, although it is difficult to “grow” large numbers of nanotubes while maintaining consistency in their properties. Helping such developments along is the aim of yet another STARnet institution, the .
The sheer variety of approaches getting serious funding shows how earnestly the problem is being taken – and also how difficult it will be to nominate a successor to existing technologies. “You always have to ask the same question: can you put 2 billion of these things on a chip, and can you scale them down further, because that’s what provides value,” says Austin. Semiconductor plants cost billions of dollars and have to be reconfigured at huge costs each time the industry moves down a transistor size. That sort of shake-up will only make sense if there is a compelling selling point – and mass demand – for the devices that emerge from new chip designs.
And the past decade might show us that speed isn’t everything. “Why aren’t consumers more angry that their computer isn’t getting any faster?” asks Cain. One answer is that, to an extent, we have been content to trade in speed for functionality. The ability to put billions of transistors on a chip has given us amazing computing power that we can hold in our hands, but recently it’s become more about faster and bigger memory, longer battery life and gizmos such as cameras and speakers.
As our world becomes more and more wired, the efficiency of a computer’s interaction with other systems, for example our own bodies, will become more crucial to its overall performance. That leads Austin to wonder whether it might be time to say goodbye to Moore’s law as the prime yardstick of computing progress. That wouldn’t be the end of computing, just the end of a particular type of computing where miniaturisation and more power delivers more value.
Then again, predicting the future of technology is a notoriously fraught business. Few people half a century ago could have predicted what the progression of Moore’s law has already brought, or what we would use all that computing power for. Perhaps 50 years from now we will have cracked the problem of quantum computing, which fully harnesses the weird correlations of quantum physics to deliver monumental low-power speed. Or perhaps we will have computers that work more like brains, processing data not in digital, but in analogue form; or electronics so efficient that they can run on sunlight or even moonlight, says .
While STARnet and projects like it search to extend the success we have already known, perhaps the most likely driver of future progress is something we can’t yet anticipate. “It will never end,” Theis says. “We’ve explored a tiny fraction in terms of what is possible for computing.”
“We’ve only explored a tiny fraction of what is possible for computing”
This article appeared in print under the headline “Faster!”
