Usually, artificial intelligence news is about software: chatbots, AI artists, or even data analysis tools. All those applications need powerful computing hardware to run on, and that hardware uses a lot of energy. All of the websites we use every day rely on massive datacentres filled with thousands of computers to function.
As the demand for computing power grows, and current technologies approach their limits, researchers are turning to the natural world and the human brain for inspiration. A new generation of hardware might be the key to faster, more energy efficient, and truly intelligent applications. Scientists are also harnessing the power of light, both in existing datacentres and in the lab, to transmit information and make computations.
The ESMH talked to three experts: Dr. Juan Pablo Carbajal, a researcher at the Eastern Switzerland University of Applied Sciences (OST), Andrea Rocchetto, the CEO of Ephos, a company manufacturing photonic chips, and Dr. Francesco Da Ros from the Technical University of Denmark, a scientist studying the intersection of photonics and machine learning. Their insights revealed that faster and more energy efficient hardware is already here, and with more research, it might change how we compute things forever.
Photonics: computing using light
Dr. Da Ros describes photonics as “a field in which physics and engineering come together to understand the natural phenomena related to light, and its interaction with matter and other particles.” The name derives from photon, the particle of light – think of photography. The name is analogous to electronics: machines that use electricity (the flow of electrons).
Andrea Rocchetto, the CEO of Ephos, describes the advantage of computing with photons: “Photons can carry multiple signals simultaneously using different wavelengths of light (a technique called wavelength division multiplexing). This means a single optical fiber can transmit many independent streams of data at once, dramatically increasing the information capacity compared to electronic systems.” – Read the full interview with Andrea Rocchetto
Rocchetto’s company, Ephos, uses both advantages of photonics to improve the performance of already hyper-optimised datacentres.
Da Ros connects these advantages to the peculiarities of machine learning.
Dr. Francesco Da Ros from the Technical University of Denmark: “ML algorithms require a higher level of data transfer between memory and processing units than traditional computing algorithms, and this bottleneck causes slower computations and higher energy consumption.”
Read the full interview with Francesco Da Ros
Photonic chips could meet these demands, though Rocchetto points out that “we’re not yet there in terms of commercial applications” when it comes to dedicated machine learning hardware. Instead, he points out that currently “photonics can help to increase the speed and reduce the energy cost of datacentres that are used to train machine learning models.”
Memristors: thinking with your “brain”
Alongside photonics, another area of research called neuromorphic computing is reconsidering how computing hardware is built. The core idea is to take “a model of how neurons work, and implements those neurons in hardware,” says Dr. Juan Pablo Carbajal. Carbajal has studied a type of device called a memristor, or “memory resistor,” that can behave in a way that mimics neurons in our own brains.
Why is this helpful?
Carbajal notes that computers as we use them today are designed to be general: the type of program you can run on a given machine is purposefully unrelated to the hardware inside it. But this generality is not natural.
Dr. Juan Pablo Carbajal from Eastern Switzerland University of Applied Sciences (OST): “Nature is able to solve problems. Our brain runs on a very specific hardware to do what we do in our life. Our mind and our brain and our body evolved all together, so they are completely entangled.”
Read the full interview with Juan Pablo Carbajal
Carbajal envisions a future where some computers are more specialised, form following function. Then, we will have devices whose “energy consumptions are similar to their biological counterparts… We’ll be able to do the same things we are doing today, much more efficiently.” Memristors are “very good for specialized hardware… You don’t have a machine that can do anything, you have a machine that can do one thing very well.”
The question is: If memristor-based chips function so differently from the generalist computers we use today, how will we adapt the machine learning and AI tools we already use? Carbajal says “the software side is where we expect to have a paradigm shift.” A lot of new, specialised code will have to be developed to act as a bridge between the programs we already have and the new generation of chips. “This is not something that’s 100% new,” past advancements in hardware have also created software challenges.
A more brain-like (or, in the parlance of the field, neuromorphic) chip might do more than just let us run our current code more efficiently. Carbajal points out that what we today call AI is not a novel technology: “we are not doing anything different from what people before were calling machine learning, from what people before were calling function fitting and before people called data regression.” The scale of the tasks tackled by “AI” has grown, but the mathematical steps of training an AI are conceptually the same as they have been for years. But, by changing the hardware, we could “get closer [in terms of intelligence] to the hardware we see in nature.”
Research to reality
Innovations in computing hardware now exist at different stages of development. Some, like Ephos’ photonic hardware for datacentres, are “already a reality” according to Rocchetto. The same goes for certain types of neuromorphic technologies, such as “cameras [and] auditory systems” that are already in use according to Carbajal. Rocchetto points out that the largest technology companies are conducting their own research into photonics: “Google has been working on replacing its datacentre network switches with photonic switches, and has been doing that for the last 10 years.” It is clear that the incentives are there to take these innovations out of the lab and into practice.
Da Ros notes that actually implementing machine learning on photonic chips is very much an “active research field with several interesting methods reported, however, none can yet rival the tools at the disposal of ML on digital hardware.” Rocchetto echoes this idea, stating that while his company isn’t actively pursuing it right now, the “long term goal is doing actual computation in a photonic chip.”
Carbajal raises concerns about how such research is supported. He says “it’s not also easy to get funding to investigate these kind of questions” when describing the fundamental work needed to turn neuromorphic chips into a viable replacement for today’s technologies. Asked what support the field of photonics needed, Da Ros and Rocchetto agreed: “Access to foundries (manufacturing facilities for chips) and effective collaboration with computer scientists and electronic hardware experts are perhaps the two major current challenges,” Da Ros replied. “Photonics would significantly benefit from more funding in photonic foundries,” Rocchetto echoed. Foundries refer to the facilities where new chips are made – an added challenge of photonics being that they require “non-standard materials for a clean room” according to Rocchetto, and thus investment in new manufacturing facilities would enable faster development of new photonic tech.
The opportunity for Europe
While the research needed is complex, Europe seems uniquely well poised to develop these new technologies. Da Ros said that “the European Chips Act provides good support,” referring to a new piece of EU legislation designed to support manufacturing. However, he also commented on the lack of trained interdisciplinary researchers able to take up the challenging work of realising new photonic systems.
Rocchetto also praised Europe as a hub of photonics innovation, saying that “Europe has a long history of strength in photonics. Innovation in photonics is also pretty widely distributed across the continent. For example, Italy is an extremely strong place for photonics, and that’s why Ephos is based there.” He also linked photonics to other novel technologies, including quantum computing, noting that “photonic information processing is needed in quantum tech, there is absolutely no future of quantum technologies without the ability to manipulate and process photons.” This interconnectedness highlights the need for interdisciplinary training, but also interdisciplinary funding.
Carbajal shared his belief that Europe is the “last bastion of what a multi-objective society can look like,” returning to the idea that progress in computing research is not the result of pursuing one goal alone, but rather the confluence of careful work in many fields. He warned, however, that pressure on academics to publish frequently “is weakening some values that […] are core values of science.” But, he returned to his belief that “Europe probably [is] one of the few regions where we are trying to have more than just one interest in focus.”
Developing new types of computing tools requires many different perspectives. Rocchetto noted the importance of materials research to determine the best composition for new types of chips, and Da Ros and Carbajal both pointed to fundamental mathematical understanding needed to develop new types of algorithms. Progress happens at the intersection of all these fields, and perhaps the European research sector is well structured to succeed in such an interdisciplinary venture.
So what?
Why should Europeans, or for that matter anyone, care about research into new computing and AI hardware? As demand grows for applications (especially AI applications), so too does the demand for energy. More efficient and faster hardware could massively reduce the energy footprint of AI. Hardware designed for specific purposes could also simply outperform our existing computing capabilities, unlocking further advancements across science and industry. The experts agree – this research is important, and its benefits are needed if we want sustainable growth: both from the environmental and competitive perspectives.
Useful link:
• What if Europe championed new AI hardware?
Related content:
• Making light work – an interview with Andrea Rocchetto, CEO of Ephos
• New steps in photonic computing – an interview with Dr. Francesco Da Ros
• Neural computing, paradigm shifts, and Europe – an interview with Dr. Juan Pablo Carbajal