ClimateTech: Powering NextGen Data Centers With Light

Photonics is coming to the rescue

Jacques-A. Gerber
BEYONDINSTITUTE

--

AI electricity demand forecast (Credit: Lightmatter)

Data centers — buildings full of computers that run all these applications you use on a daily basis from your preferred electronic devices, including what you’re using right now as you’re reading this, as well as more and more business and government applications — are estimated to consume about one percent of our global electricity use and generate about one percent of energy-related emissions. This is comparable to the airline industry as a whole, which is responsible for over two percent of energy-related emissions, according to the International Energy Agency.

While the airline industry might grow moderately as more people get back on planes, data centers are expected to grow much more significantly. With new AI algorithms (e.g. chatGPT) and new applications (e.g. metaverse), demand for computing power is exploding. So much so that the Semiconductor Industry Association predicted computers would need more energy than we produce by 2040!

Data centers need a lot of energy not just to power electric circuits and processors but also to cool them down. Processors generate so much heat that they would melt if they were not cooled down. This requires air conditioning and sometimes liquid cooling. As a result, cooling represents up to 55% of a data center’s energy consumption. More energy-efficient processors, cooling systems, and energy management software have helped improve data center energy efficiency but demand keeps growing alarmingly faster. One way or another, the computing industry needs a radical change.

But what if we could get away from heat-producing electric systems and replace them with light and optics? Well, that is the field of photonics, and it is a revolution that is happening now.

Computing systems need to do three things: compute, which is what processors (CPU, GPU, TPU…) do; store data, which is what memory and drives (RAM, SSD…) do; and exchange data through wires and fiber optics.

The processors are the brains of computing systems. They’re also the components that need the most energy and generate the most heat. And we need a lot more of them to keep up with the demand for new AI algorithms and applications. In other words, Moore’s law is not sustainable.

And that’s where photonic processors come to the rescue! Lightmatter has started shipping a processor that computes using light. This provides higher performance for a fraction of the energy needed by traditional processors. Indeed, the company claims it can run 10 times faster than NVIDIA — the leading AI processors company — for 90% less energy.

Lightmatter Envise™ photonic processor (Credit: SPIE)

The founding team comes from MIT where they conducted the initial research leading to their breakthrough technology. Founded in 2017 with the support of Google Ventures, among others, Lightmatter has attracted top executive talent from leading companies such as Google and Intel.

This solution is especially well suited for AI and High Performance Computing, which require the deployment of a very large number of processors. For instance, they won a $4.8 million project by IARPA to develop self-driving car technology with Harvard University and Boston University. They have also been recognized as one of the 100 most promising AI companies by CB Insights in 2022 and for making the world a better place on Fast Company’s World Changing Ideas.

In an interview conducted by Tech Brew in July 2021, Nick Harris — Lightmatter CEO — explained that replacing processors in 100 data centers with its product would save enough energy to power 62 million homes on an annual basis!

Let’s now talk about data transfer. You probably have heard about fiber optics. They transport data with light over thousands of miles between data centers, hubs, and buildings around the world. If you’re among the lucky ones, they might even reach your home. These are also used inside data centers to connect racks of machines as well as one machine to another. You might also be familiar with relatively short optical cables that connect a CD player to an amplifier. That is to say, we have been using fiber optics for transporting data for a long time. But once they connect to a machine, a sophisticated component called an optical transceiver converts data from light to electronics and vice versa. Inside the machine, everything is electronic, data is transported on copper wires all to and between processors, memories, and drives. All this electronic circuitry is slow and generates heat.

But what if we could replace all these copper wires with fiber optics? Well, that is exactly what a company called Ayar Labs is doing. After a decade of research involving teams from UC Berkeley and MIT, they made a breakthrough. They succeeded in making a very tiny component that can transmit information between electromagnetic waves and light waves as transceivers do, but at a much smaller scale. Not only does it work, but it can be manufactured at scale and at a competitive price.

Graphics showing Ayar Labs photonic components connecting two processors (Credit: Ayar Labs)

This technology can increase bandwidth by 1000x at 10x lower power. If a data center was fully equipped with this technology, it could reduce its energy consumption by up to 50%.

Ayar Labs is now working with companies like HP Enterprise, Intel, and Nvidia to build their next-generation computing systems with their technology.

After processors and connectors, remains the question of storage. The bad news is there is no optical solution for storage, given that it is practically impossible to store light. The good news is that storage components do not come close to consuming as much energy as processors and copper wires.

Photonics has a long history in computer science but is only now coming to fruition. By replacing electronic processors and wires with optical solutions, we will drastically reduce data centers’ energy consumption while formidably increasing their processing power.

For a more technical review, check this excellent article published in SPIE. and written by Mark Wolverton

Disclosure: my company — Agave Partners Management I, LLC — manages a private fund that is a shareholder in both Lightmatter, Inc. and Ayar Labs, Inc. I, therefore, have an indirect interest in both companies.

Lightmatter™ and Lightmatter Envise™ are trademarks of Lightmatter, Inc. Ayar Labs™ Ayar Labs SuperNova™ and Ayar Labs TeraPHY™ are trademarks of Ayar Labs, Inc.

--

--