New hardware harnesses the computing efficiency of the brain

Supercomputers are the backbone of a broad range of scientific disciplines: modelling the weather, disease diagnosis, AI chatbots, the list goes on.

The problem is, they’re often extremely vast machines requiring huge amounts of electricity to run. The world’s most powerful supercomputer, the HP Enterprise Frontier, requires 22.7 Megawatts. That’s roughly the same amount required to power 17,000 homes.

It’s for this reason that scientists are looking at alternative avenues to achieve the amount of compute to complete their data intensive research projects. Neuromorphic engineers believe the answer could reside in the amazing efficiency of our brains, and 2024 is set to be a big year for the field.

James Tytko caught up with Domenico Vicinanza, Associate Professor of Intelligent Systems and Data Science, Anglia Ruskin University, to find out more...

Domenico - The idea comes from the observation that even the most powerful supercomputer in the world cannot compete with the competing power that our brain has. We can truly multitask; we can walk and talk and speak on the phone, observe and avoid obstacles. We can truly do all these things at the same time. That's fascinating. There is no supercomputer in the world that can do that in the way we human beings can. A computer can simulate that multitasking by doing a little bit of each single action, each single task, and then quickly moving to the next one. Those characteristics are related to the different computing architecture that our brain has. That's the reason why scientists were inspired by the different approach of our brain. And our brain doesn't overheat. It only weighs 1.3/1.4 kilos, consumes 20 watts, like a light bulb in our fridge, and it can do the equivalent - well, it's actually more powerful in terms of basic computation per second - than the most powerful supercomputer. So there must be something that we are missing; some lessons to learn in some way.

James - So that's the idea but how does one achieve that, practically? What aspect of our brain can scientists try to emulate inside technology?

Domenico - To answer this question, we need to think a little bit about how our brain is made. We know a normal, traditional computer (I'm oversimplifying here) but it has one computing unit, one memory, a bus - a highway where data are going to and from - that's it. This is the way a normal computer works. A brain doesn't have that. There are no specialised areas in the brain that are just dedicated to computing or just dedicated to memory. We can think about our brain as a really intricate network of very simple computing and memory elements that are together. Where there is computation power, there is also memory in our brain and that is absolutely brilliant. The other very important thing is that all these areas are talking to each other through trillions of connections.

Domenico - That means when we are learning something new, we learn a new skill like pitching a ball, for example, or learning a new language, what we are doing, we are not just putting few lines of code somewhere in our brain that will be retrieved, we are reshaping our brain. The connections between all the neurons that will describe and serve that specific purpose, building new connections, that kind of flexibility is fascinating. It's inspiring. That is one of the things that scientists are trying to reproduce: having the machine equivalent, mechanical equivalent of our neurons and our synapses in a single chip. The idea is to focus on connecting them with, in the case of the new supercomputers - they're going to be large - we are talking about hundreds of trillions of connections. That is the power.

James - You've prefaced it there. This field of neuromorphic engineering, neuromorphic computing has reached something of a watershed moment or is going to this year, isn't it? Can you tell me about Deep South and whether this is where neuromorphic computing really starts to kick on now?

Domenico - I think yes. Neuromorphic computing is starting to kick off now in a proper way, in a more disruptive way than it did in the past. It's not a new idea, first of all. Neuroscientists had ideas about how the brain actually works and how important the connections are between neurons and how relevant this intricate network is that can reconfigure for many, many years. The first ideas about neuromorphic computings were in the 80s. Of course, the technology was not good enough to create something that can have the computing power of even a small brain. Today we can. For different reasons; technology miniaturisation, engineering, we can do that. Create something that can mimic the number of neurons and number of connections of our brain. How quickly and efficiently they can reconfigure, they can actually layer things, that's a completely different matter.

Domenico - We don't know yet. The first one that will have the same scale of our brain is called Deep South and it will be launched, switched on in April, 2024 this year at the University of Sydney. What makes it special compared to its predecessors? What is different is the scale. We are reaching brain scale. It's fascinating to actually have something that, in principle, can emulate with a big...

James - Asterisk.

Domenico - Exactly. Emulate our brain. And what is interesting is that because it's inspired by the same architecture, it means that we can actually run some models on this, not just having more efficient computing. That's one of the aims. So we want to hopefully reduce the power. We want to have more computing power per watts that we're spending, but also this can open a completely new field which is how we can actually understand a bit better the way the brain works, how brain medications work on our brain, what is the best way of alerting something or the best way of treating some neurodegenerative, for example, disease. By knowing how the model that is, in this case, a mechanical electrical model of our brain can respond to the same stimulus.