Lightmatter have just released a new kind of computer one that based on light and this is a big deal for the entire industry let me explain why today computing demand is growing faster than silicon chips can keep up with to get more performance chip makers nowadays just throwing more silicon at the problem double the area double the RAM double the cost and it's been working for now but there is a catch because the rule of the game in semiconductors is that you pay per area per silicon area used and the costs nowadays are skyrocketing nowadays a single GPU costs way more than your rent one thing is clear we can't double down on silicon we have to rethink how we compute well if we think about it at the data center scale it comes down to 3 main aspects first of all compute interconnect and memory let's start with compute for the past decade the main engine behind AI from transformer models to reasoning models has been accelerated matrix math GPUs pushed it TPUs refined it and then ASICs squeezed out every drop of efficiency and now the spotlight is shifting to computing with light but why in fact we don't know much about dark matter but light matters a lot it's the source of energy growth and time and now it's not just powering the life on Earth but also computing you've likely heard this idea that light-based computers are faster because light travels way faster than electrons well it might be partially true but it misses the real point let's take a regular chip AMD or NVIDIA GPU for example it's built off hundreds of billions of transistors those tiny little switches that are constantly turning on and off to perform computation and it's getting even more interesting here because actually in digital chips every time we want to switch from 0 to 1 or from 1 to 0 we have to stop the data take time to either charge or discharge a capacitor think of it like filling a tiny basket with electric charge just to flip a switch this takes time and now imagine doing it billions of times per second this is where the real slowdown is coming from and this is exactly where photonic computers shine because those are analog chips not digital and this makes the whole difference in light-based computers we are using light waves and light doesn't have to stop to charge up there is no capacitance like with silicon this means we can process data on the fly without any delay for switching and that's why photonic chips are so much faster now one very interesting thing to understand about light-based computers that those are governed by Maxwell's Equations and Maxwell's Equations are linear and this has a huge effect on what these kind of computers can actually do it turns out that at the core of modern AI workloads are actually additions and multiplications and those are linear operations and that's exactly what photonic computers excel at now let's put a spotlight on it and see what happens let's say you want to do a matrix multiply accumulate 128 by 128 so if you do it on the Lightmatter photonic processor you get result back in roughly 200ps well how this compares to a conventional GPU on a conventional GPU for this you would roughly need 100 cycles and if we take 1ns per cycle this is roughly 100ns what I'm saying is that the photonic processor can do the whole job under 1ns which is roughly very roughly 100-1,000 times faster so you see how much faster we can compute when we don't actually have to stop the data another very interesting property of light is that it operates at much higher frequency here we are in terahertz range compared to the gigahertz range in electronics this practically means that we can compute much more data simultaneously and by using different colors of light we can compute lots of data in parallel and this practically means that we can achieve massive parallel computing without spending more area or more power just think about it this is mind-blowing despite this glowing interest in photonics there is a catch or two first of all analog chips are super efficient but this comes at a cost of precision until now analog chips have never achieved the precision that we actually need you don't want your banking transactions to run on a light-based computer because so far they were nowhere near the precision of the digital chips now Lightmatter finally solved this with their new photonic chip in a very elegant way in fact they managed to achieve a precision that is very close to precision of 32bit digital chip and I was lucky to get an opportunity to discuss it with co-founder and CEO of Lightmatter Nick Harris we've built the first alternative computing system that doesn't use transistors that's able to run economically valuable and useful workloads things that you would actually want to run what we were able to do is we built a photonic computer that can play Atari video games it can run Transformers it can run large language models and I think that's a historically significant milestone you had to prove that an alternative form of computer like a photonic computer could run these workloads accurately as accurately as a digital computer and that's what we did now let's have a closer look at the photonic engine it's essentially an accelerator designed to accelerate linear algebra which boils down to adds and multiplies if we break it down this new photonic computer includes photonic tensor cores and electronic chips integrated vertically via high speed links first of all there are two electronic control chips that are on the top of the chip package their goal is to communicate with the photonic tensor cores and then there are also four photonic engines which are 3D stacked underneath you see all nonlinear math is actually offloaded to the digital chip while all the heavy math like multiply accumulate is happening in the light domain so in total inside there are six chips in a single package and overall it's 50 billion transistors that coordinate 1 million photonic devices we will dive more into the details how it works later on but looking at the high level a digital chip sends a request to the photonic engine and then in roughly 200ps gets the result back now one picosecond is 1 trillion of a second so this happening very fast now think how of all of this is synchronized it's a rocket science if we have a closer look into the photonic engine here you can see the photonic core the wave guides through which light propagates during the computation and here is a close shot at the electronic part of the chip which is of course a bit less interesting because we can't see much due to the metal layers which are hiding all the beauty of the circuits underneath now let's get laser focused on how actually computing with light works let's say we have some data an image and it's described by a vector which represents pixels in the image and the values are between 0 and 1 describing how much red blue or green it is first we map this vector into the optical domain then the light travels through optical devices and get multiplied by weight using so-called Mach-Zehnder interferometer (MZI) and multiplying here actually means turning a number turning a weight in how bright the light actually gets then the light arrives to the end points which are all connected to a single electrical wire so in this way signals get summed along that wire meaning additions happens naturally and this is the true beauty the true power of light you know no delays no clock cycles it's all happening effortlessly just pure speed let me know what you think in the comments next we will explore the performance of this chip and what it means for the future of computing but before that have have you ever wondered how much of your personal information is circulating around the web your name your home address phone number and even information about your family members could be floating around online this happens because data brokers collect and sell your personal information without you ever realizing it and this exposes you to risks of data breaches and personal security and this is where Incogni the sponsor of today's episode comes in Incogni helps you regain control by removing your personal information from the databases that data brokers rely on i use it myself and you will be surprised how simple it is you sign up authorize Incogni to act on your behalf and they send data protection low compliance requests to these companies forcing them to remove your information from their databases and the best part you can track every step of the progress in real time right from your dashboard as someone who values privacy a lot I highly recommend you to try out Incogni it's a simple way to reduce unwanted spam and keep your data off the grid use my code INTECH at the link below to get 60% off an annual plan thank you Incogni for sponsoring this episode analog chips have been long time dismissed as too imprecise of course if you can't get math right computer is not useful and now Lightmatter for the first time got it right and I asked Nick to explain their elegant solution well we have a number format called ABFP16 and what we do is we assign to a block of numbers a scale so we factor out a number we save that in the digital processor and then we query the photonic tensor core to do those adds and multiplies we get the result which is a vector and we multiply it by the scale function again but that's just one part of the equation they're using other tricks as well for example they are over amplifying small signals so neural networks are not losing critical bits in our terms it's called LSB's least significant bits to simplify it think of it like zooming in on the most important math numbers close to zero and it works for the first time light-based chip achieved the precision close to digital chips and it's not just a demo but a real functional chip well what we witnessed over the last couple of years that AI is moving in the other direction from 16 to 8 bits and now to 4 bits now 4bit format is becoming a new standard because it's allows us to reduce compute and memory requirements and this shift represents a huge opportunity for photonics because it turns out each time we drop precision efficiency increase exponentially photonic engines are crazy efficient at low precision if you look at just the tensor core itself so just the math engine that's operating at a few hundred tops per watt which is extremely high energy efficiency it's also performing at that efficiency at a very high throughput so that's a tricky point to be at when you want to go fast with a digital computer you spend energy and it's nonlinear so going twice as fast costs more than twice the amount of energy in many cases with this system you're in this quadrant of it's very fast and it's efficient so that's a very interesting spot to be in of course real world systems have other components that eat into efficiency but it turns out there is a lot of room how we can further optimize it for example now Lightmatter is using just one color for computation but they can easily increase it to 16 or 32 and then reuse all the components to perform massive parallel computations just think about it let's say they go from 1 color to 16 and they immediately have 16 times higher throughput or higher computational density if you will without significantly increasing area this could power the future of intelligence let me know what you think in the comments and if you're enjoying this episode remember to subscribe to the channel this makes me and my team very happy now all of this sounds brilliant but we know that when it comes to exotic computing approaches they either take decades or even never escape the lab take analog computers that are based on some sort of a resistive device such chips have huge potential but are very unflexible and struggle with running AI models as you can't just run it out of the box it requires translation of the models and even additional training while the new Lightmatter processor can already run Deepmind's Atari and nano GPT which is a reduced version with 100 million parameters and it doesn't require any translation of the models or any additional pre-training that's great but remember at the very beginning of the video we discussed that photonic chips are governed by Maxwell's Equations and there is a second catch because they can easily accelerate linear operations but unfortunately they can't manage logic the challenge with light is that photons don't interact beams pass through each other like ghosts in order to make them to feel each other you need exotic nonlinear materials and deep difficult physics and that interaction in fact is the essence of logic light doesn't know how to play this game that's why in the future we are likely to see photonic engines accelerating linear math and probably financial trading but it's unlikely that they will run Linux or Windows at least not anytime soon then there is one more fundamental problem with photonics as actually one of you pointed out in the comments it's really an honor that such bright smart people are watching my videos thank you for your comments so when it comes to computing let's say you want to invent a new computing paradigm from scratch what we need to do we need to be able to manipulate signals like add them multiply them and then we must be able to remember intermediate results so we can use it for the further computations or to act on this result now when it comes to photonic we can perfectly manage the first two but there is no storage available do you remember we discussed those capacitances which slowing down digital signals this is the way the intermediate results stored in digital chips and this does not exist in photonics what typically happens in a photonic chip we will convert this light signal into a digital one so back to ones and zeros and this is a slow part and also it drains a lot of power this means computations that don't require the memory truly shine in photonics but we still have to figure out the memory part and to be honest this is not the problem that everyone is focusing on right now we are on this channel is a couple of steps ahead of the rest of the world now everyone is figuring out how to efficiently link those large GPU clusters together and when doing so interconnects matter because in modern AI workloads no single chip does the job alone here thousands of GPUs work in parallel and they constantly exchange data even nano second delays in data exchange between GPUs have a huge impact on the time it takes to train an AI model if we manage to solve that we can release new AI models way faster and not only that there is a new class of models so-called reasoning models like DeepSeek R1 or so-called Deep Research models those are very accurate but it takes them 10 minutes to generate a solution for you so if we can solve the interconnect bottleneck and connect more GPUs together efficiently we can reduce this response time from 10 minutes to let's say 10 seconds this would be cool the solution is to replace copper that is currently being used to link up racks with photonic interconnect and Lightmatter is solving it with their Passage product so this is the big opportunity for photonics and we're building the fastest photonic engines in the world we announced M1000 at our event a couple weeks ago M1000 is 114 terabit per second in a single optical engine we've built platforms for customers that are 60 TB per second in a single optical engine we announced L200 which is our standalone general purpose IO tile for GPUs and for switches 64 TB per second so Lightmatter is really the bleeding edge on how fast these systems can be we're about 8 to 10x faster than any of the companies that are out there people are announcing 8 TB and 6 TB we're at 64 and as soon as this interconnect bottleneck is solved everyone will start looking more into the improving the efficiency of computing and here photonics can enable the next big leap clearly the future is optical optics will be everywhere let me know your opinion in the comments now please let this video to see the light share it with your friends colleagues and on social media i really appreciate your support finally if you're obsessed with light as much as I do check out this episode where I explain new photonic chip from NVIDIA and basically the main trend which is happening in the industry right now must watch thank you for your support and I will see you in the next episode ciao
2025-05-02 12:41