COMPUTER & SUPERCOMPUTER: THE FUTURE OF THE PC
Will we still have personal computers in the future, and if so what forms might they take? Ian Evenden investigates.
Will we still have PCs in the future, and what forms might they take?
Twenty years ago, when PCs were universally beige, and APC’s staff had a lot more hair, we probably would have nodded in agreement if someone had told us that the desktop’s days were numbered. We’d used laptops, we’d heard about Moore’s law, we knew things didn’t get much better than Windows 95. It stood to reason that portable computers would take over, and that 17-inch CRT we could barely carry up the stairs would soon be on its way to the dump. That’s partly correct, as we thank any passing deity for flatscreens, but the desktop PC is still with us, and is really the only way to go if you want no-compromise computing power, be that for gaming, 3D content creation or 4K video editing.
Laptops are fine, and our phones and tablets have come a phenomenally long way, but there’s always something that holds them back, be it insufficient cooling, limited RAM, the drawbacks of mobile processors or, in the case of touchscreen devices, an unintuitive user interface.
So the desktop endures, looking much the same as it did in the days of the 80286, only rotated 90° into tower cases, and with more fans and lights, and a whopping great water block. Will it still be purring away on our desks in another 20 years, though?
One future route is already happening with the advent of cloud computing. Our PCs could become dumb terminals, akin to the thin clients of old, tapping into the huge processing and storage potential of data centres using high-speed wireless connections. They could all become touchscreen devices, small enough to tuck away in our pockets, then unfolding to gigantic size like some Tony Stark creation. They could be implanted into us, like something out of Iain M Banks’s Culture novels or an episode of Black Mirror, connecting us to a central hub, and with a display like that of a smartphone constantly projected into our vision.
Or maybe they’ll just stay the same, doubling in power every two years, and needing a new GPU even sooner. We went looking for answers.
UP CLOSE AND PERSONAL
The term ‘PC’ has evolved from meaning any type of personal computer to specifically those involving X86 processors and running Windows. Elsewhere, X86 + Unix = either Mac or Linux, ARM + Unix = iOS or Android, and ARM + Windows = discontinued. But those ARM chips are what’s got Samsung poised to overtake Intel as the world’s biggest processor manufacturer, and isn’t the smartphone in your pocket as much of a personal computer as the one on your desk?
Many developments in future PC components are extensions of things we’re already familiar with: greater efficiency, using less power, and increased parallelism. A Ryzen 1800X may have eight cores, but a GTX 1080 Ti has 3,585. Yet while AMD’s silicon is general-purpose, Nvidia’s is specialised for graphics computations. Bringing this kind of parallel computing power to the mainstream, via something such as Nvidia’s CUDA, Microsoft’s DirectCompute or other GPGPU programing languages, is a step in the direction of what’s predicted for the immediate future of our PCs. Then there’s 3D chips, such as those seen in Xpoint or the vertical stacks of AMD’s HBM — these are about more efficient use of space, as well as, in the case of Xpoint, making use of a new approach to operating a memory chip. These technologies are here now, even if they’re not widely available. Future technologies will take these innovations, and turn up the volume.
Looking ahead, by 2035, we should have cracked thermodynamically reversible computing — that is, logical operations that can be run backward from their result, because nothing is destroyed in their operation, and they don’t increase entropy. This sounds a bit crazy — and is really only the tip of a whole iceberg of crazy — so we spoke to Professor Robin Hanson of George Mason University in Virginia, who is also a research associate at the Future of Humanity Institute of Oxford University in the UK, in an attempt to understand.
“When we run an engine, entropy is increasing,” he says, prompting a trip to Google immediately. Entropy, it seems, is the amount of energy in a system that’s wasted, unavailable to do work. The professor continues: “But if we want to reduce how much entropy increases, the slower we make the engine go, and the closer we can come to what’s called a reversible process, where you could have made it go backward and got back to the original state. That’s also true for computers: Almost all logic gate operations take two bits in and they send one bit out, therefore implicitly erasing one bit, and increasing entropy. Typically, we’re erasing far more than one bit per gate operation, but that number has been declining over time.”
This is where the year 2035 comes in — if you plot the number of bits erased by transistor operations on a graph, and continue the line into the future, it’s a mere 18 years until it reaches one. “At that point,” Prof Hanson continues, “we could keep reducing it, but we’ll have to switch to reversible computing, where we have two bits in and two bits out, with nothing erased. Once we switch to reversible gates and reversible computers, then we can continue to reduce the amount of entropy per gate operation, but it will be because we run the gate more slowly. If you take the same gate, but take twice the time to do the gate operation, then it erases half as many bits. So when hardware gets cheaper, and energy gets cheaper, you will spend half of it on having more hardware, and half of it on running things more slowly.”
So what we are looking at are lots of slow, parallel processors, rather than the small groups of screamingly fast cores we see today. But that’s not all: It will change the upgrade cycles, and what we budget for. “At that point, the rate at which technology improves, or hardware costs go down, is half as fast as it has been so far,” says Prof Hanson. “Up until recently, energy hasn’t really been the main cost — that’s been the hardware itself.
“The Tunnel FET looks like a MOSFET, and you can build a lot of the same circuitry with them.” Dr Kirsten Moselund
And so, every time the hardware is capable of running twice as fast, the computers are twice as fast. But with reversible computing, every time the hardware gets four times as fast, instead of running the computers four times faster, we will instead have twice as many of them running at half the speed. This means Moore’s law will slow down by a factor of two.”
Moore’s law, which predicts a doubling of computing capacity over a period of two years, is already starting to break down in areas such as the speed of gate operations, but holds strong in the amount of energy used per operation. Many scientists are looking at ways to improve the rate of speed increase, and decrease the energy usage, through the use of new materials in chip design, and by changing the design of the transistors themselves. IBM researchers in the US are attempting to press nanotechnology, such as carbon nanotubes, into service, while in Switzerland, Big Blue is testing out other naturally occurring elements, known as III-V materials.
BACK TO SCHOOL
Picture the periodic table thumbtacked to the wall of your science class in high school. Silicon sits in the fourth column, with columns III and V to either side. Many elements from both sides, when made into compounds, form very stable chemical bonds, and are semiconductors, like silicon, which means that although, in their natural state, they don’t allow electricity to flow through them, they can be ‘doped’ with another element to allow it, and this can be done in a controllable way. A lot of this is still exploratory science, but you may already own a piece of technology based on this idea — the laser diodes inside Blu-ray drives are made from gallium nitride — and hopefully more should come to fruition before 2035.
“The speed at which you can turn on a MOSFET is inherently limited,” says Dr Kirsten Moselund from IBM Research. This is the breakdown in Moore’s law mentioned above — the MOSFET, or metal oxide semiconductor field effect transistor, is the type of transistor most commonly used in a silicon chip, first patented back in 1925. “It doesn’t matter what you make it out of, it’s a physical limitation,” she continues. “As we scale down our devices, we also want to scale down voltages, but it’s very hard to scale down power beyond 60mV — you start to get into trouble. There are lots of people looking into this, because it’s beyond all the technical difficulties of scaling, it’s something physical. And scaling down the voltages is probably the most important parameter for energy efficiency.”
The search for a way around this has led Dr Moselund and her team to Tunnel FETs, a type of transistor that exploits the ability of electrons to tunnel through a barrier if that barrier is thin enough. This is quantum mechanics in action, making use of a strange property of electrons — that they can be either waves or particles.
“The Tunnel FET looks like a MOSFET, and you can build a lot of the same circuitry with them,” says Dr Moselund. “But as it operates on a different principle, in theory you don’t have the same limitations.” Indeed, a Tunnel FET can operate on only about half the power of a MOSFET — in simulations, at least. These transistors are made from, you guessed it, those III-V materials — indium arsenide and gallium antimonide are common choices. “III-Vs have a lot of really nice benefits,” Moselund continues. “There are lots of them, and they have different properties, but what’s generally good about them is many have very high electron mobilities [how quickly an electron, and therefore a current, can move through them], so you can trade speed for lower power, and many are optically active, so you can make lasers out of them,
which you can’t do with silicon.” Lasers are obviously great, but modern electronic devices have another problem — they leak. This is why they get hot and drain their batteries when you’re not using them. The gates in the transistors don’t shut completely, allowing small amounts of power to trickle through like a leaky faucet. Tunnel FETs and III-V materials have a greater ability to turn themselves off all the way, decreasing leakage and power loss.
The new materials don’t completely replace existing ones; they’re integrated into the silicon in a way that boosts its electron mobility, to increase performance at 7nm and smaller, meaning existing manufacturing processes can still be used. IBM, GlobalFoundries, and Samsung debuted a silicon wafer etched with a 5nm process earlier this year. The III-V materials are grown as crystals on a silicon substrate, then a process known as ‘epitaxy’ deposits more material on top, forming structures such as nanowires and junctions, and even stacking them on top of one another. You can also mix up the recipe — what Moselund calls ‘tuning’ — blending, say, 50% arsenic with 20% gallium and 30% indium, with a specific use in mind.
It’s not just the structure of microchips that will change; the way computers are put together and treated is also in for a revolution. Professor Hanson believes that computers will one day be able to simulate and go far beyond the capabilities of the human brain. He calls these emulations ‘Ems’ and his concept is not a little thoughtprovoking: When computing is advanced enough, it will out-compete humanity.
“A natural result is that humans will have to retire,” says Hanson. “They’re just not competitive. They could still work, they just can’t earn much money that way. Collectively, humans get very rich very fast, that is, they own almost all the capital in this world, and if the economy doubles almost every month, human wealth doubles every month. For individual humans who don’t own any wealth, that zero keeps doubling to zero, and those people are at risk of starving unless they acquire some insurance, assets or sharing arrangements. But collectively humans get very rich very fast.
“The Ems, however, they do not get rich,” he says. “Their population quickly expands and wages stay at subsistence levels, but they are mostly OK with that — subsistence wages have been the usual case in human history. They’re earning enough to pay the power bill, the cooling bill, the hardware rental, the communication line bill... which will presumably come in one big package.”
This is treading rather closer to philosophy and economic theory than we’re used to, so in an attempt to drag it back to more familiar terms of reference, we ask Hanson whether he thinks there’s a future for the desktop PC. And he does. Kind of.
“A lot of people will continue to do office work at desks — it’s kind of comfortable to sit in a chair,” he says. “And they will want something around them to act as the interface to the computer they work with. Now, whether the computer is just sitting on the wall, or in their hands, or on the desk, that’s much harder to say. They could have a box next to them or a server down the hall — it hardly matters from the point of view of them interacting with the computer. I like to use the same computer at home and in the office, so I prefer a laptop, but the incentive to use a laptop will get less the more reliable cloud services get.”
Hanson goes on to imagine cities that glow red hot because they’re occupied by the hardware and cooling needed to run Ems, with humans banished to more habitable parts of the globe, but the inference is clear — in the near term, PCs will continue to get more powerful, integrating new materials and structures into their designs. But as the cloud gets more important and reliable, and communication links get faster, a transition to something like today’s supercomputer model could occur, with virtual computers held in the cloud, and eventually virtual workers there, too. It’s safe for now, but the days of the desktop PC could ultimately be numbered. Sorry.
A Tunnel FET can operate on about half the power of a traditional MOSFET transistor.
Nvidia’s Xavier SOC uses 512 GPU and 8 CPU cores to be the brain of an autonomous car.
Also from Nvidia, the Tesla GPU introduces additional parallelism for AI use, and more.