Lightelligence

YICHEN SHEN: My name is Yichen Shen, I'm the CEO and Founder of Lightelligence. Lightelligence is a company spinning out of MIT trying to commercialize my PhD research results, which is using light to accelerate computation. So I grew up in China with passion from when I was seven, eight years old in physics and mathematics and I came to US when I was 20. in my PhD, I mainly focus on nanoscale photonic phenomena.

So how light interact with materials and interact with light when they're at very small scale. I told my advisor that I really want to look into this intersection between photonics and AI, and at the time I realized that computation is actually one of the biggest enabler for modern stage artificial intelligence and faster computing hardware is going to be able to make an algorithm that smarter and be able to make decisions quicker and also by paying attention to the internal algorithm of AI, I realized that it's actually very friendly for optics to carry out such algorithm.

So after I graduated from MIT I spent one more year at MIT as a postdoc to wrap this whole project up and at the time, we also attended several business plan contest including have a president challenge at MIT 100K. Mainly trying to validate the commercial value of this technology and trying to make a decision whether what's the next step to do with this, so luckily we win first prize in both competition and that basically confirms my understanding of how useful this technology is and that also helped us raise our next round and the following round and gather the team together and start a company.

Our founding team is our MIT students, our alumni. So two of my friends, one obtain his PhD from ECS and the other team member was in the same lab that I was from since then we have now have about 40 people in the company. We have very well-established industry veterans seeing semiconductors including a senior fellow from AMD and then we also have very talented students or alumni from MIT Harvard, Stanford, Columbia University's in semiconductor, in photonics in artificial intelligence and software. So now we have a very well rounded team and they work seamlessly together to bring this technology to a reality.

[MUSIC PLAYING]

SPEAKER: Our technology was utilizing the same fabrication platform as any traditional semiconductor chips. So basically we make small microscopic scale or nanoscale component had devices on the silicon platform but instead of making transistors which is typically what an actual chip does, we make optical, waveguide optical from mirrors and those optical components on the chip at micrometer scale. So what optical devices really breena's which electronics cannot is, for example in a traditional electronic chip in order to do any arithmetic operations, you need to combine tens or hundreds of logic gates to make a single modification.

Well, in optical domain, we do all the computation with physics instead with logic. So it was basically, we precisely control how the light interact with each other inside the chip and when the light is interacting with each other they automatically does the computation. For us it's just like propagating through the chip and when it's probably getting through the chip, it interfere with each other and the nature of such interference does the mathematics that we want to do. So that's how our optical computing chips actually work.

And the advantage of this over electronics is where light is propagating in the chip, it barely generates any heat. So as you can imagine, optical fibers across the Pacific Ocean-- the light propagating through thousands of kilometers without losing much power. While an electronic chip, instead, every single logic gate transistor's flowkey is generating heat. And the main problem for current Moore's law is to slow down, and for the electronic chips not to be able to speed up is because of the heating problem, while optical computing chip doesn't have this problem.

And this advantage enabled much lower power consumption obviously much higher computational throughput at a higher Clark and also much lower latency to compete complete one task. We think it's a much more effective investment into optical computing because it does have a lot more potential than what this electronic can have.

[MUSIC PLAYING]

SPEAKER: So one of the biggest challenge from lab to real world when I was a PhD, I only need this device to work once. I'll have 100 devices. So I can make 100 devices try each of them 100 times and I just need one device to work for one time. But when you really want to make it a product, not only you need 99 out of 100 devices to work 99 out of 100 times at the same time, you have to care about cost, you have to care about reliability, you have to care about how much product you can make in the year, and all this commercial where are the killer apps and beachhead markets.

So that's a lot more, it's a lot harder. Yes we are at early stage, I would say we're still at R&D stage for our technology but industry really is a very I would say vague word because as I mentioned there are so many different applications and each different applications have a different tolerance for what is industry ready I would say. For us, this I would say standard curve that we have to define what is the early adopters which they are industry ready is probably are at like a working pilot demo that's a complete system PCI plugging, running an algorithm that they care about that's already industry ready.

So these are early adopters. Mainstream adopters are probably like those data center and so on they need reliability tests, they need all the software to be ready, and then also like PCI, powers, and everything has been proven and also early adopters have said yes. I have already said yes to this technology, so that's probably industry ready for them. So I would say for early adopters we are looking at industry ready in one or two years, given the current customers that we're talking to for mainstream I would target like a 3, 5 years time period for us to be industry ready.

So I think next year we'll get to the point when we will have a hardware that we have direct 80% to 90% of the technical challenges, for a lot of optical computing to be commercially viable product. So that was on the technical development side and at the same time I'm looking at getting into more like a commercial relationship with a few pilot customers that we're talking to nowadays.

[MUSIC PLAYING]

YICHEN SHEN: The funding team of Lightelligence is-- I would say the team that really came up with this idea and invented this technology of using optics to do AI computation, we wear at least one of the first team that start looking into this space and we think we are also the first team in the world that I'm aware of that actually build a complete system of optical computing hardware.

So after we incoporate the company, what we does is that we build a dedicated chip that's mainly designed for optical computing purposes and that chip was everything is integrated-- all the optical devices are integrated on that chip including on chip modulator, on chip detector, which was not previously on our MIT prototype and the chip is able to run at gigahertz [INAUDIBLE] which is a million times faster than what was in the MIT demo.

And what is more when we're at MIT all the other components like the electronic components and the lasers and so on they are off the shelf and they are not integrated in the system. So if you really look at the setup it's almost as big as half of a room or at least half of an optical table, the whole system but for this time we managed to integrate everything into a laptop space and make it portable and really looks like a standalone computing system.

Now, in the past year more than a year we have already started developing much more close to a commercial product like demo and so what we are scaling up is we are going to scale up the size of the matrix to 100 of 100 sized matrix. We are going to scale up the clock rate and also we are going to integrate all the electronic component to the electronic chip to make it even more compact. So that's I would say the scaling up roadmap for us.

Well and then talking about on the other side of the competition, I wouldn't say that we only have competitor in optical space, our main competitor is actually digital electronic. When the first car company come out, they're not going to say our competitor is another car company. We are going to say our their competitor is forces or like other, what is mainly used at the time. So actually more competitor that's doing up to a computer is good for us at this stage because we have a louder voice to the community and to bring up the whole ecosystem for optical computing.

[MUSIC PLAYING]

YICHEN SHEN: So at this stage, we're mainly working with a few tier one, I would say, potential customers in a few different applications, including data centers, OEMs, and other niche applications where they have a real need for high performance hardware. So the areas that we are interested in and are digging into nowadays are data centers, self-driving cars, civilian cameras, finance, crypto, and also health care.

First I think ILP bring us a lot of connections to industry like we're mainly looking at tier one potential customers. So ILP give us a lot of connections in this way. And at the same time, it's very fortunate to have the mentors and also the opportunities to get into those ILP events to talk to people and to [? networking ?] with the partners from ILP and industry partners. I think that's what we benefit a lot from that.

[MUSIC PLAYING]

YICHEN SHEN: Over the past few years, the industry have witnessed a rapid rise of artificial intelligence. If you look down the road for the next few decades, it is obvious that the amount of data and the intelligence decisions that people have to make will continue increasing exponentially. But it is unfortunate that the computational power is probably not going to catch up at the same pace as the need. At Lightelligence, we use light instead of traditional electronics as the fuel to power our computing hardware. With Lightelligence product, we are able to achieve the same amount of computation in AI tasks at a fraction of time with a fraction of power.

[MUSIC PLAYING]