More speed, less energy

Roughly speaking, the human brain can perform 100,000 times more operations per second than a single computer chip, using around a tenth of the amount of energy. That makes the brain the indisputable lodestar for the next generation of computer hardware, according to MESA+ professor Wilfred van der Wiel. He and his colleagues, including professor of Programmable Nanosystems Prof. Dr. Ir. Hajo Broersma, are betting on nanomaterials as the key to a brain-like system.

‘Don’t get me wrong, conventional computing can still take us a long way,’ says Van der Wiel, who is professor of NanoElectronics at the University of Twente. ‘Players like Facebook, Google, Microsoft and Twitter are getting better and better at artificial intelligence (AI), such as tagging faces, recognizing individual voices and tracking online behaviour. For simulating neural networks, people have used supercomputers and special hardware, such as graphics processor units (GPUs) field-programmable gate arrays (FPGAs)  and tensor processing units (TPUs). But these systems are all still based on conventional computer hardware: they require a logical design and operate on a linear basis, performing in sequence each separate computation required for solving a given problem. Their big limitations are time and energy.’

 The chips are down

Add to that the facts that chips today are no longer getting faster, while demand for advanced computing capacity is soaring, and you’ll understand why all eyes in the computer world are on nano scientists like Van der Wiel and Brinkman. Van der Wiel explains, ‘The chip has reached its maximum speed. We’re only gaining capacity by putting more transistors on a single chip and dividing the chips across multiple cores: more is better. Oh, and we’re reducing their size, of course: currently the minimum chip size is around 14 nanometres. We may be able to cut that down to ten or even seven. But a transistor the size of two atoms is just not possible. Moore’s law – which says the number of transistors in a dense integrated circuit doubles approximately every two years – is about to be outdated.’

What conventional computing can and cannot do

Does this mean conventional computation is going the same place Moore’s Law is going? Not by a long shot, says Van der Wiel. ‘There are many things conventional computers can already do a lot better than the human brain can, for example simple arithmetic. And the number of applications for that kind of computation is growing all the time, of course. But when it comes to fuzzier, less clearly defined tasks – tasks that require complex simultaneous operations, like pattern recognition, simulating complex pharmaceutical compounds or machine learning – the brain-like approach is incomparable. And demand for that type of neuromorphic, or brain-like computing is also growing rapidly. It could open up an entirely new era for computing.’

 Ten, fifteen years, maybe a lot less

 ‘No one can tell how long it will take for us to build a genuinely neuromorphic system,’ adds Van der Wiel. ‘Ten, fifteen years – maybe a lot less. But the first steps, like the proof of principle for a nanoscale network carrying out logic operations, have been taken. Companies like IBM and Google are also very active in this field. It is a matter of time before we reach the next breakthrough.’

While no one quite knows how the first brain-like system will be built, or how it will operate, Van der Wiel and his nanoelectronics research group at MESA+ are betting on nanomaterials. ‘What is new and different about our approach is that we are looking for a way to build a system from the bottom-up. So rather than designing a network that can perform pre-determined logical computations, as in conventional computer hardware, we have taken a designless system as our starting point. Instead of having a central processing unit (CPU) carrying out pre-set instructions, our system – which is really nothing more than a heap of two hundred gold nano balls surrounded by a bunch of electrodes – is ‘blank’, so to speak. The functions it can carry out depend on the properties of the nanomaterial and how they evolve under varying electric current conditions. What we found is that by playing around with the current, we can create any logic circuit we want. We just don’t know yet how it happens. The trick is to recognize the right configurations so that they can be reproduced.’

Solving the power problem

On a global scale, Van der Wiel and colleagues are ate the very forefront in developing the kind of reconfigurable, neuromorphic materials with controllable and dynamically adaptable electric or magnetic properties that emulate the neurons and synapses in the human brain – and that any soon will pave the way for the world’s first nano network. Key goals in this pursuit are speed and energy consumption reduction, says Van der Wiel. ‘IT already uses a lot of energy and the rate at which it is spreading further and deeper into every area of life on our planet suggests we have a problem to solve there. Nano computers could be the answer. The human brain can perform computations vastly more complex than anything a state-of-art computer can do today using just 10 to 20 Watts. That is an exciting fact.’

By Stephen Teeuwen for MESA+