NcodiN: Shaping the Future of Optical Data Communications between Chiplets

As chips pack increasingly more transistors per square millimeter, monolithic systems-on-chip face significant challenges. A single faulty die can spell disaster for an entire chip, leading to costly manufacturing setbacks.

To tackle this problem, many chipmakers are looking into disaggregating chips into chiplets—individual dies specialized for a specific task, reusable, and easy to manufacture. The real challenge then becomes ensuring seamless data communications between those chiplets.

NcodiN is on a mission to redefine chiplet-to-chiplet communications with its optical interposers. Founded by Francesco Manegatti, Bruno Garbin, and Fabrice Raineri in the spring of 2023, it develops semiconductor nanolasers that can be attached to a single chiplet to beam information between them. The technology promises a breakthrough for the on-chip convergence of electronics and photonics, paving the way to high-bandwidth, efficient, and low-latency data communications between chiplets.

Learn more about the future of optical data communications between chiplets from our interview with the co-founder and CEO, Francesco Manegatti: 

Why Did You Start NcodiN?

The story began during my Ph.D. I was in my third year and started hearing more and more about startups coming up in France, particularly quantum startups. At the time, I was working on nanolasers and figured that, even beyond academic research, they could address a problem in the real world––how to communicate data optically. 

I went to my supervisor and suggested that I continue working on this project even after I finished my doctorate—not just as a postdoc doing academic research, but to start a business. We could also benefit from several government grants and lots of support from the French National Center for Scientific Research (CNRS). Funny enough, that very same day, my supervisor had wanted to discuss this with me in the afternoon, so we were already on the same page. 

We applied for a special grant by CNRS, called the pre-maturation project, which allowed us to move forward with the R&D for a year and a half while also exploring the business opportunity. During that time, we managed not only to get great scientific results and fabricate nanolaser devices ourselves but also standardize the fabrication and get better reproducibility from one batch to the other. We saw this had the potential to become viable for industrial manufacturing. That’s when we eventually decided to spin out and found NcodiN in the spring of 2023. 

How Do Your Optical Interposers Work?

We use light as the carrier of information to make chips communicate with each other. In particular, we transform an electronic signal to an optical one, literally by switching on and off a nanolaser, and allow for faster, smoother, and more energy-efficient data communications. 

Our nanolasers work like a diode. When we inject electrons, they enter the so-called active region of our devices and stimulate the emission of photons. We use a special 2D material for the active region that traps these electrons in certain points, called quantum wells, like a ball getting stuck in an actual well. However, because our wells are tiny, even on the nanoscale, they are subject to quantum physics, which states that electrons trapped in a quantum well can only assume discrete energy states.

Think of it like boxes on a storage shelf. They can only be on one floor in the rack, but not in between; otherwise, they would fall to the next lower floor. So, their potential energy can only assume discrete values as determined by the position of the floors. 

Similarly, electrons in a quantum well can have only certain discrete energies. When they fall from a higher energy level to a lower one, they emit a photon whose energy corresponds to the energy difference between them. By tuning the discrete energy levels in a quantum well, we can tune the energy and hence the wavelength of the emitted photons. 

Building a nanolaser is then a matter of stimulating coordinated photon emissions from all those quantum wells and building a cavity whose geometry amplifies particular wavelengths through reflections and eventually emits a laser beam. The receiver relies on the same principles but in reverse. By attaching emitters and receivers to different chiplets, we can communicate data between them.

We benchmark our optical interposer on three metrics:
1) bandwidth: how much data we can transfer per second, where we reach petabits per second transfer rates; 

2) latency: how much time it takes to move from one chiplet to another, where only the speed of light in silicon is our constraint, resulting in about 0.1 nanoseconds per centimeter;

3) efficiency: how much energy we need, where we get down to 10s of femtojoule per bit. 

For comparison, electronic interposers, i.e., putting a copper wire from one chiplet to another, can only optimize for either bandwidth or latency: 

To minimize latency, you need to put multiple parallel copper lanes so signals don’t get into each other’s way on a single wire. You’re limited due to capacitive interactions between the wires to a few millimeters. If you want to connect hundreds of chiplets and even those a few centimeters apart, you must pass through all the other chiplets in between. It’s like passing through tolls every few kilometers on a highway. 

To maximize bandwidth, you need to serialize your data, i.e., use additional circuitry for serialization and deserialization, which makes you lose time and increases latency. 

With photonics, you send data at the speed of light and use different wavelengths within the same channel to transfer multiple data streams and achieve much higher bandwidth. Also, one of our unique selling propositions is scalability: our nanolaser can be attached to chiplets in a plug-and-play fashion, so our technology can scale from a few to hundreds and eventually thousands of chiplets. 

Other startups explore optical interposers using an external laser, where the electro-optical conversion comes from micromodulators that encode data by modulating the laser beam as it passes from one chiplet to another. This requires a larger, monolithic integration overall and the external laser to power the entire architecture. With each optical element, you lose a bit of intensity, making it harder and harder to power everything the more chiplets there are.

Thus, you need increasingly stronger external lasers, which take up space you don’t have. The only way forward for complex architectures involving hundreds and thousands of chiplets is to integrate multiple laser sources on-chip. That’s why we use many but small lasers from the start and place them only where needed.

One of the main challenges is thermal stability: Even though processors in data centers typically have a cooling down system attached so that the overall system temperature goes down to 45 degrees Celsius, the temperature often increases on-chip locally up to 70 degrees Celsius and may peak even beyond 100 degrees Celsius. They must be able to withstand these temperatures and function reliably even with temperature variations for at least three to five years, as they are integrated into data centers that are rarely upgraded.

We have demonstrated the first fundamental component of our optical interposer, the nanolaser, showing that we can meet the needed optical power output with low electrical power input and produce them with good yield and reliability. Next, we’re developing the receiving end, the photodetector, which will be another major milestone towards an end-to-end proof of concept. 

How Did You Evaluate Your Startup Idea?

It was a matter of interviewing many industrial players, executives, and engineers to understand which roadmap we have to meet and the main challenges they face. It’s really a lot about talking to people, understanding their pain points, and only then pitching our solution. And talking to clients of our clients, as the whole semiconductor industry is intertwined and talks to each other—and there are just about a dozen or two chipmakers that we’re targeting as major customers. 

The industry is moving away from monolithic systems-on-chip, which have reached extreme sizes that make them hard and costly to manufacture. Instead, we see the disaggregation of chips into chiplets, so chips can still become more powerful with an increasing number of chiplets. As individual chiplets are a lot smaller, they can be manufactured with extremely high yield, reused across various chips, and allow for a faster time to market. Optical interposers will then allow for fast and efficient communication between the chiplets. 

What Advice Would You Give Fellow Deep Tech Founders?

When I look around me, I see many PhDs that could start exploring entrepreneurship—but then many don’t because they fear the unknown, trying something entirely different from what they’re used to in academia. Academia is about characterizing hero devices and writing publications about them. To build a business, you need to get to manufacturing at scale. And you also need to deal with people outside of academia and explain to them in simple terms what you’re doing. It’s a matter of practice to communicate clearly and understandably.

Ultimately, entrepreneurship is a matter of practicing and gathering experience that you can’t get from reading books or the success stories of others. Take your phone, call people, learn about their pain points, pitch your idea, and get their feedback. And you can’t do everything alone—find the right people to join your mission. Finding the right person for the right spot is one of the main skills of a CEO. And again, don’t be afraid to talk to people! 

Further Reading

Francesco Manegatti, Grand Prix I-PhD – Innovation Award for his project NCODIN developed at C2N – Learn more about Francesco’s research project from C2N’s news announcement

Comment la deeptech Ncodin veut booster les communications entre puces grâce à des photons émis par des nanolasers – Zettascale computing systems need ultra-fast and low-power interconnects such as NcodiNs to ensure the scalability of future multi-chiplets-based HPC and AI processors. Featured by the Industrie & Technologies news platform (French only)