AAA Data centres turn to photonics to cope with the demands of AI

Data centres turn to photonics to cope with the demands of AI

Speed, light
Photo by Matthew Brodeur on Unsplash
Photo by Matthew Brodeur on Unsplash

The world is entering a new age of hyper-scale data use. The number of devices that connect to the internet is expected to grow 30% to 38bn by 2025, according to IEEE. Average internet traffic per user is expected to double every three years. And the advent of artificial intelligence and machine learning will add massive pressure on bandwidth capabilities.

Data centres are having to be redesigned to cope with this explosion of data and photonics — using wavelengths of light rather than electrons to transfer data — is becoming a key part of that data centre revamp.

Fibre optics is already widely used to transmit data at faster speeds and with greater bandwidth than the flow of electrons through electric cables — telecoms networks have for a long time been moving from copper to fibre. But copper wires and integrated circuit chips are still the main workhorses in the server racks inside data centres.

That is changing, as data centres move to a new model, where components such as the central processing unit (CPU), storage and memory – are separated so that they can work more efficiently. In traditional data centres, servers are loaded up with a fixed ratio of components like CPU and memory. But, as workloads become faster, those devices are not used evenly, and sometimes will even be dormant for periods of time. In order to run data fast enough between these separated areas, data centres are looking to photonics.

“We’ve started to see this emerging trend where certain workloads that would have traditionally been done with electrons are offloaded into an optical mode,” says Owen Lozman, managing director of M Ventures, the corporate venture arm of German biopharmaceutical company Merck. 

Optics is making its way into the rack of data centres and even into the semiconductor chip itself, to cope with the massive amount of compute power required for new technologies such as language models that produce bodies of text as part of natural language processing.   

A new crop of startups is emerging in the field of photonics to address this data centre market.

Owen Lozman, managing director, M Ventures

M Ventures has invested in early-stage photonics startup Polaris Electro-Optics. The US company is designing a next generation modulator, a component in the integrated circuit that inscribes information to prepare it for transmission as an optical signal.

Polaris is pioneering the use of a new material to replace silicon in the manufacture of modulators. Silicon modulators are “running against their limits in terms of how fast they can go and how energy efficient they can be,” says Jason Sickler, co-founder and chief executive of Polaris Electro-Optics.

The startup has found that a type of liquid crystal added to its modulator design can make it faster and more energy efficient than pure silicon.  The technology for the liquid crystal was discovered at the University of Colorado Boulder in 2020.  Polaris is exclusively licensing the patents. The company is raising a seed round that it plans to use to increase the material’s performance and then demonstrate a high-speed device, says Sickler.

M Ventures’ parent company, Merck, is already familiar with the liquid crystal that Polaris works with because it uses the same type of material in the manufacture of LED screens for TVs, iPhones and other electronic products as part of its display business.

“Our display business has been a very strong business for us over the years. That business is becoming quite mature, and we are looking to move into new areas where we can see a lot of growth,” says Lozman.

Another startup in the field of photonics is Lightelligence, a US company that makes photonic integrated chips and optical interconnect products that connect compute devices optically.

Like Polaris, Lightelligence targets the data centre market. Its customers are hyperscalers – large internet companies that provide cloud computing and data storage services. Examples of hyperscalers are Amazon, Facebook, Google and Microsoft. These companies run their own customised computer systems designed to accelerate artificial intelligence and machine learning applications.

The startup also targets enterprise data centres that buy ready-made equipment through suppliers such as HP and Dell.

The company is much later stage than Polaris. It has raised $220m to date and expects to manufacture its products in 2024. The names of its investors are undisclosed.

“It is pretty inevitable that you are going to have to make date centres significantly more efficient in the use of their resources. The value of that is going to be enormous.”

Jason Sickler, co-founder, Polaris Electro-Optics

“It is pretty inevitable that you are going to have to make date centres significantly more efficient in the use of their resources. The value of that is going to be enormous,” says Sickler.

Jason Sickler, co-founder, Polaris Electro-Optics

In disaggregated data centres, resources are pooled together and optimised, rather than co-located in the same server. Designers also do not need to guess what is the right ratio of CPU and memory to build inside a server. This removes any potential dormancy when one computing component is running more than the other. This optimisation of components in the data centre is the only way that AI language models can grow, given the massive amounts of data processing required.

In this new generation of data centre, servers are connected by optical interconnects. The technology is still early stage, although semiconductor makers have started to focus on the sector.

“Optical interconnects have really gained lots of ground and interest in recent years. Ten years ago, there was nobody from the big semiconductor companies officially talking about photonics. But now there is acknowledgement that photonics will be an important backbone for various industries, notably now in data centres facing skyrocketing demands due to the rapid evolvement of AI,” says Tobias Egle, an analyst at M Ventures. “As a result, semiconductor companies now have dedicated fabs, some of which are relatively small and some more research based.”

“Ten years ago, there was nobody from the big semiconductor companies officially talking about photonics. But now there is acknowledgement that photonics will be an important backbone for various industries.”

Tobias Egle, analyst at M Ventures

The lack of a photonics equipment manufacturing ecosystem is a challenge for startups. The technology is still very early stage. “Customers want a mature ecosystem. You would think being a frontrunner is great, but you also need an ecosystem. A lot of data centres don’t want to rely on one vendor because they have requirements to switch you out,” says Spencer Powers, co-founder and president of Lightelligence.

The lack of investment in photonics hardware means it will take time for the sector to evolve. The semiconductor industry, on the other hand, has had decades and billions of dollars of investment. “It is super hard to compete with that,” says Lozman.

But he adds that the limits of electrical integrated circuits will ensure photonics becomes a critical part of the semiconductor manufacturing sector. “It’s only when people see what we’re up against with the scaling limits of electronic components, that people will realise we need another technology to address these bottlenecks,” says Lozman.     

By Kim Moore

Kim Moore is the editor of Global University Venturing and deputy editor of Global Corporate Venturing and produces video for the website.