Artificial intelligence is transforming the way computing systems process and move data. As AI models become larger and more complex, the demand for faster communication between processors, memory, and other components is creating new challenges for data centers.
Traditional computing architectures have made significant progress in processing power, particularly with the rise of general-purpose GPUs. However, as GPUs increasingly work in parallel, the movement of massive amounts of data between them has become a critical bottleneck.
In this episode of the Analytics Insight Podcast, Rohin Y, Founder and CEO of LightSpeed Photonics, discusses what led him to start the company, why AI is forcing the industry to rethink data center infrastructure, and how photonics can help overcome the limitations of traditional electronic interconnects.
Ans: We actually started seeing the symptoms of the problems almost six or seven years ago. And now that the time has come for the applications, most people have started recognizing it. However, when people say "photonics," I think they may not recognize it immediately. However, when you say fiber, you instantly recognize it with something like geofiber or ethyl fiber in India. Basically, fiber to the home is what people know to be the fastest because light is the fastest there is.
So hence the name Lightspeed Photonics. Literally, you can transfer data at the speed of light, and that is the only means to communicate the huge amount of data that is required in the future of data centers or anything.
Ans: Data centers, they're basically just storing data. It was a huge infrastructure setup put together to store data and serve people over the internet, and that required only minimal management of data. Yes, it has gotten more and more complex with what they used to call, in the 2010s, big data analytics.
So what shifted was, of course, the general-purpose GPU computing, as we call it.CPUs are sequential in nature. So they basically process data one bit, one byte at a time, one instruction at a time, whereas GPUs can handle data in a much more parallel manner. So, massively parallel amounts of data streams have to be shared across these GPUs. This is one of the key reasons why the interconnect started becoming a bottleneck.
Ans: So if you think from what we were discussing how the data should be exchanged, and networking power growing higher- you have to also see that historically we faced, the industry faced what we used to call a compute wall.
Meaning we were not able to process data fast enough, and obviously that is the thing that general-purpose GPU computing solved in a major way that they were able to compute fast. And we went beyond the compute wall, and then we hit what we call the memory wall.
Ans: So this is a very interesting trend. Overall, in silicon, we know how to scale; we know how to produce in such huge volumes. It's literally the cheapest thing you can buy, right? The transistor is what got, when it was invented, a few dollars' novelty to its nano paisa today, right? It's the cheapest thing available in the universe, right? You can literally buy those transistors. So you know how to scale with silicon. We have all the foundries and the fabs.
So when silicon photonics investment started, the key goal was to build on the infrastructure that knows how to scale up silicon-based wafers into compute chips and use it for photonics.
Ans: I think India is in the best position to take advantage of photonics in general because we, not just India, but many other countries in the world, have had a difficult time catching up to what CMOS silicon technologies have grown. So, in practice, Taiwan has a monopoly in that domain.
It doesn't require us to build huge fabs, but we can do it with relatively minimal OSATs and advanced packaging with the heterogeneous integration I was talking about and fiber attach. Many surrounding industries can grow. In fact, we are working with many partners and trying to build that ecosystem here to make this more meaningful and much, much larger.
To know more, listen to the full podcast.