Mike Davis, Intel Labs: “We’re approaching the limits of what basic computing can do” | Technologies

The constant increase in data traffic (according to DE-CIX, 22% more last year compared to 2022) and the new computing needs of artificial intelligence mean that systems…

Continuous growth in data traffic (up 22% last year compared to 2022, according to DE-CIX) and new artificial intelligence computing demands are pushing traditional systems to their limits. New formulas are needed, and quantum computing is not yet an alternative. Electronics company Intel is at the forefront of developing neuromorphic systems, a marriage of biology and technology that aims to mimic the way humans process information. Together with her in this race for more efficient and effective processing they run IBM, Qualcomm and research centers such as California Institute of Technology (Caltech)where Carver Mead gave birth to this concept, MIT (Massachusetts Institute of Technology)He Max Planck Institute for Neurobiology in Germany and Stanford University.

This month, Intel announced the world’s largest neuromorphic system: Hala Pointwith 1150 million technological neurons and 1152 processors (chips) Loihi 2, which consume a maximum of 2600 W and have processing power equivalent to the brain of an owl. Research published in IEEE study It provides greater efficiency and performance than systems based on central processing units (CPUs) and graphics units (GPUs), the conventional computing engines.

Dallas-born Mike Davis, who turns 48 in July, is director of neuromorphic computing at Intel Labs and is responsible for the latest advances that will shape the immediate future of computing.

Ask. What is the neuromorphic system?

Reply. It’s a computer-aided design inspired by modern understanding of how the brain works, meaning it beats traditional architecture by seven or eight decades. From a basic point of view, we are trying to understand the principles of modern neuroscience in order to apply them to chips and systems to create something that works and processes information more like the way the brain works.

TO. How it works?

R. If you open the system, the chips, you can see very striking differences in the sense that there is no memory; All elements of computing, processing and memory are integrated with each other. Our Hala Point system, for example, is a three-dimensional network of chips, like a brain, and everything communicates with everything, just as one neuron communicates through the brain with another set of connected neurons. In a traditional system, the memory is located next to the processor, and the processor constantly reads data from the memory.

Hala Point is a three-dimensional network of chips, like a brain, and everything communicates with everything, just as a neuron communicates through the brain with another set of connected neurons.

TO. Is this model necessary because we are approaching the limits of traditional computing?

R. Great progress is being made in the field of artificial intelligence and deep learning. This is very exciting, but it is difficult to imagine how these research trends will continue when the growth in computing requirements for these AI models is growing at an exponential rate, that is, much faster than the development of manufacturing. We are approaching the limits of what this basic computing architecture can do. Additionally, if you just look at the energy efficiency of these traditional AI chips and systems compared to the brain, the difference is many orders of magnitude. It’s not so much that traditional computer architectures aren’t capable of delivering big advances in computing and artificial intelligence, but that we’re striving for greater functionality by having computers that act like brains, and do so very efficiently.

TO. Is energy efficiency the main benefit?

R. This is one of the main ones. There is a big difference in the efficiency of the brain and the efficiency of traditional computing. But brain-inspired neuromorphic architectures could also provide performance gains. We consider GPUs to be incredibly high-performance devices, but really only if you have a very large size and a lot of processing data available on disk or near the CPU for reading. But if the data comes from sensors, cameras or video in real time, then the efficiency and power of traditional architectures is much lower. This is where neuromorphic architecture can really provide huge gains in speed and efficiency.

TO. Does artificial intelligence need a neuromorphic system to grow?

R. We think so. But we are at the research level. Today it is unclear how to implement this for commercial purposes. There are still many problems to be solved related to software (programming), algorithms. Many traditional approaches do not work initially Hardware (commands) are neuromorphic because it is a different approach to programming. We believe this is the right path to achieve the energy efficiency and performance improvements we need for these types of workloads, but the question remains open.

question. Will we see a neuromorphic chip in a computer or mobile phone?

R. I think so, it’s a matter of time. It won’t be next year, but the technology will mature and be implemented in edge computing (processing data close to its source to improve speed and efficiency), mobile phones, autonomous vehicles, drones or laptops. Our Hala Point, designed for the data center, is a box the size of a large microwave oven. But if we look at nature, we see that there are brains of all sizes. Insects are very impressive, even on a small scale. And then, of course, the human brain. We are investigating on both fronts. We believe that commercialization will begin in edge computing

but there is a need to continue to push and conduct research on a larger scale.

We will be able to see these systems in data centers in five years

TO. When will they be?

R. This is difficult to predict because there are still open questions in the research. We will be able to see these systems in data centers in five years. We also see a future in anything that needs to be battery powered, as the energy savings that a neuromorphic system can offer are extremely important. There are also less obvious applications, such as wireless base stations for telephone infrastructure. We are working with Ericsson to optimize communication channels.

TO. Is computing using neuromorphic systems complementary to quantum computing?

R. I think they complement each other in a way, even though they are very different. Quantum computing is looking for innovation in the production of physical devices and trying to scale. What it offers is very new and impressive, but it’s unclear what Quant’s programming model will be, when it will scale, and what workloads it will support. The neuromorphic computing available today is very good for AI workloads. But there is an intersection in the field of quantum and neuromorphic applications, and that’s where it’s interesting to think about solving complex optimization problems and allowing people to experiment, prototype, and learn how to program these types of systems.

Implanting neuromorphic chips into the brain is a natural application for these systems because, as an architecture that behaves like neurons, it naturally speaks the language of our brains.

TO. Can we see the neuromorphic systems installed in our brains?

R. There are some researchers interested in neuroprosthetics, the application of neuromorphic computing, which means trying to correct problems or pathologies in the brain where there has been some loss of function and regaining control of your body. The research is in its early stages, but I think in the long term this will be a very natural application of neuromorphic computing because, being an architecture that behaves like neurons, it will naturally speak the language of our brains.

TO. Available systems, what age brain are they equivalent to?

R. It is similar in number of neurons to an owl’s brain. But if you look at the area of ​​the brain where most higher-order intelligence occurs, it is equivalent to a capuchin monkey. Many of us in this area of ​​research have the human brain in mind as a kind of vision for the scale of the system we would like to build. But we’re not trying to get there too quickly. We must build it well and know how to make it useful. That’s why this system remains a research tool so we can continue to experiment.

TO. In what specific cases are these systems most effective?

R. When finding the best path on a map, we see up to 50x speedup compared to the best traditional solvers. In terms of energy, levels that are 1000 times more efficient are achieved.

TO. Can Europe take advantage of this new line to regain its footing in the race for chips on which it depends on other continents?

R. If we look to the future, we will need a lot of innovation over the long term to match nature’s size and efficiency, which remains incredibly impressive. We still have a long way to go, and to do that we need innovation in manufacturing. New devices and new memory technologies are needed to assimilate them into the brain. There is currently no geographic region with an advantage in this area, so this is an opportunity. High technology always means innovation, and nothing remains static. There is a need for new achievements, and it is unknown where they may come from.

You can follow El Pais Technology V Facebook And X or register here to receive our weekly newsletter.