- 3 months ago
AI is advancing at a breathtaking pace but its hunger for computing power and energy is growing even faster. As traditional compute approaches its physical limits, the industry faces a fundamental question: What will it take to sustainably scale the next generation of AI and HPC workloads?
In this session, Michael Förtsch, CEO & founder of Q.ANT, explores the emergence of photonic computing, a radically new computing paradigm that uses light instead of electrons to process information.
From the basics of light-based processing to real-world AI workloads, this talk will examine how photonic architectures could unlock orders of magnitude gains in performance and energy efficiency and open pathways to fundamentally new classes of algorithms.
As demand for AI accelerates across every industry, photonics may become a key technology that enables scalable, sustainable computing for the decades ahead.
In this session, Michael Förtsch, CEO & founder of Q.ANT, explores the emergence of photonic computing, a radically new computing paradigm that uses light instead of electrons to process information.
From the basics of light-based processing to real-world AI workloads, this talk will examine how photonic architectures could unlock orders of magnitude gains in performance and energy efficiency and open pathways to fundamentally new classes of algorithms.
As demand for AI accelerates across every industry, photonics may become a key technology that enables scalable, sustainable computing for the decades ahead.
Category
🤖
TechTranscript
00:00.
01:00We should start now, right?
01:02So the question is, is photonic computing the answer to the energy crisis we are having
01:10in AI?
01:11Yes, full stop, and I could go.
01:14Who is aware of photonic computing?
01:19Okay, that's good.
01:21Because I thought it's worth spending a few words what it exactly is and what's the differentiator.
01:27But to walk you into the topic, I guess we all hear we like AI, right?
01:32Honestly speaking, these slides have been prepped by using one of these AI tools.
01:40And there seems to be no limit where the imagination in terms of revenue and market growth comes,
01:49neither on the infrastructure side nor on the application side.
01:52But what, in my opinion, is largely ignored in the public debate is what is the energy
01:59cost of what we're doing here.
02:01So on the right side, there is a conservative chart of the energy consumed by the data centers
02:08worldwide.
02:08And if you don't like charts, the numbers, next year, it's going to be the energy consumption
02:15of Japan.
02:16And the hard thing about the scaling exponential law is that in 2035, it's going to be the world
02:25energy if we scale with the same technology as we've been scaling the past.
02:30It's not going to happen, but then we already realize that the end of our vision at this point
02:37is not performance, it's not imagination, it's energy.
02:45Where does the energy go to?
02:48Not to the processor.
02:51It goes to the memory.
02:54Because what we've been improving on since decades is the energy efficiency and the performance
03:00by shrinking the transistors down, effectively the compute performance in terms of energy
03:06got better and better.
03:07What has not coped with that was the energy consumption of the memory.
03:11And I just listed a few comparisons here for the non-physicists.
03:16This is picojoule, so joule is an energy unit.
03:19And you see that most of the energy is consumed by devices like the HPM or the DRAM.
03:26And the minor part actually goes on the compute.
03:30So that's where the energy in the stack is flowing to.
03:35And whenever things got tricky in history, think about it.
03:39When we ran into issues with electricity, we switched into light, maybe because of bandwidth
03:48or maybe because of energy.
03:50When you look up, the great thing about light is once it's generated, it's propagating.
03:56You don't need to push it.
03:57You don't need to push it with another energy to move electrons.
04:02Electrons always want to be kicked.
04:04They always want to be energized by something.
04:07You work against the resistor.
04:09Light is just propagating.
04:10That's a very simple argument.
04:11I'm coming to the more sophisticated in a bit.
04:15And while five years ago, the large part of the compute industry, it's going to be copper, copper, copper.
04:22The reasons for that was copper is cheap.
04:26They switched.
04:28They switched within two years.
04:30And we've seen how this led to acquisitions in the photonic market.
04:35And there is no need to believe that it's going to be the last part that's going to be flipped
04:39into photonics.
04:40Because the electrical part on the processing side is as well running into issues.
04:46So this is where we are.
04:48We are at the processing side.
04:50We are not at the data transport side.
04:51And we make computers that compute with light.
04:55To walk you a bit through the stack, I know the slides are a bit technical.
04:59But believe me, at the end, I have applications that are already operating.
05:02So it's worth staying with me on these slides.
05:05When you look on such a computer, it's not a single chip.
05:08There are multiple chips.
05:09And all of them are used.
05:12There are the host processors, the CPUs.
05:15You know AMD and Intel.
05:16You know RISC-V.
05:17You know ARM.
05:18So that's the part where you run your operating system on.
05:20And it's going to be operated on that in the future as well.
05:23Because they're a perfect match.
05:25And then we switch to the coprocessing side.
05:27A GPU is obviously a coprocessor.
05:30It doesn't run the operating system.
05:31It does specific operations very fast and very good.
05:36We are a coprocessor.
05:38But in the contrast to the digital part, we are an analog part.
05:42Now, most of you in this room, I assume, have been here on the quantum discussions before.
05:46Just to also make a note, their quantum computers are great.
05:50But we are not a quantum computer.
05:52We are an analog computer that performs analog computations, not in the quantum space.
05:58So we are a coprocessor to the GPU and we are side-assisted to the CPU.
06:03What do we do?
06:05Obviously not quantum.
06:08Internally, we say we compute the hard shit.
06:12The complex stuff.
06:14We compute mainly what usually you would hate to put on a GPU.
06:20Non-linear equations.
06:21Things where usually I obviously lose parts of the audience, typically, because this is where
06:27most people stopped in math.
06:29But we need these equations.
06:30Why do we need these equations?
06:32Because the world that we want to model with AI is way too complex to just be modeled by linear
06:37equations.
06:38We do that as of today.
06:40But that's one of the limitations and that's one of the root courses why we need so much data.
06:45So this is where we position the photonic.
06:48Photonic is great.
06:49Now, why is photonics so great?
06:51And this is a qualitative chart.
06:54On photonics computing, the energy you need to compute plus and minus is the same amount
07:03of energy you need to compute a sine function or a full Fourier transformation.
07:09It doesn't matter whether you understood.
07:11It's just the degree of complexity of functions was just increased, but the energy stood exactly the same.
07:18So that's where we see the large winning space of these processors.
07:22The more complex the function gets, the stronger it gains in comparison to the CMOS world.
07:29Why is this important?
07:31Because now you can start trading.
07:34The most simple example that I'm having is if you want to model a circle and you only have linear
07:42equations,
07:44you need a lot of data points to come close to what the circle is.
07:48If you inherently have a sine function, there are two parameters and you describe the problem perfectly.
07:56It's an oversimplification given, but that's the point where I want to guide you to.
08:00So with this processor for the first time in compute history, it makes sense to start trading.
08:06Trading data complexity into function complexity, because the function out of a sudden doesn't cost you more.
08:15And that's where you start gaining energy efficiency.
08:18It's a complex argument, but I'll show you the results.
08:21In the first years, we've been focusing on pictures, because pictures are obviously complex things.
08:27And what you see here is the performance benchmarks on the individual applications over two years of bringing processors into
08:36the system.
08:37We started with digital handwriting recognition, simple tasks.
08:42It's a standard routine, so-called MNIST dataset.
08:45And over one year of improvement, we basically went to be performance equivalent to the state of the art of
08:53CMOS.
08:55Then we focused on image classifications.
08:58As of today, with the current generation, the generation two, we make two pictures a second.
09:05Everyone in CMOS would laugh now, but I guess you stop laughing next year, because then we've outperformed you.
09:11We already have benchmarked that we're making 40,000 pictures a second.
09:15In the same quality, with the same accuracy that you are used to.
09:21We did this on image segmentation, and the last one is pointing towards the AI space on the generation two.
09:27We, for the first time, also trained pictural models.
09:31And the fun part here now is that we use up to 50 times less parameters to train a neural
09:38network that, for instance, creates Van Gogh pictures.
09:42That's saving energy.
09:44Now, yet again, into the details on the image classifier, because this is where I have a systems breakdown.
09:51If you want to classify images, you would use an algorithm like the ResNet-50, meaning 50 layers of convolutional
09:58layers.
09:59And you would, per picture, need roundabout 50 megabyte in your memory.
10:05Now, let's be gentle and say we reduce the accuracy, so we go from FP16 accuracy to INT8 accuracy, then
10:12we have half of the memory needed, approximately.
10:16We need a fraction of that.
10:17So we can do this with 7 megabyte.
10:20Why?
10:20Because we use Fourier transformation and more complex stuff in the algorithm, and we do not even have to move
10:28the weights.
10:28They're living in the photonic mesh.
10:30So that's where the speedup comes from.
10:33The proven point here is we have four times less weights in the network.
10:38We have four times being faster than the state-of-the-art technology at a 10x energy reduction on the
10:46same application.
10:48So it's not just percentage that we move the needle, it's on the 10x scale here.
10:56Some of you might think, oh, God, analog computing, it might be hard to now work with these guys.
11:03It's not.
11:05On the T-Systems challenge this year, together with a partner, DaisyTuner, we demonstrated the first compiler.
11:12And what we did is we used a model that was relying on PyTorch.
11:16We didn't touch the surface code.
11:18It was exactly the same, and the compiler basically compiled it down to be operatable on a photonic engine.
11:25So you don't need to change anything of your coding habits.
11:29And that's the result.
11:31It's image segmentation in real-time identifying objects.
11:37And as a remark, it's the same recognition quality as you would have on a regular CMOS-based architecture.
11:46It's working.
11:47It's a real experiment.
11:48We brought service to Bonn, and we performed it live on stage.
11:54So it's real.
11:57Serenity, I heard it all around.
12:00Now, we can talk hours and hours about why I'm not a big fan of trying to copy what the
12:05value chain or what the semiconductor industry has built over the past 40 decades.
12:10This technology that I'm demonstrating you is entirely built in Europe.
12:15We even go down to the material level.
12:18When I founded Quant in 2018, we didn't start at the top.
12:22We started at the bottom.
12:23We built the wafers for that.
12:26We own the design and layout.
12:28We even build a pilot line.
12:30So every chip that you're seeing from Quant is built on a pilot line in Stuttgart.
12:38We own packaging, and we own the software stack.
12:42So it's the end-to-end approach.
12:45And why is this feasible?
12:47How did we, for instance, get hands on a proprietary pilot line?
12:51Photonic chips don't require 3-nanometer node technology.
12:56What we did is we found a partner who ran a 90-nanometer legacy fab, being state-of-the-art
13:04around about 1990.
13:07And that's exactly the structural width you need for photonic chips.
13:11We repurposed it, we put our processes for the manufacturing on the machines,
13:16and since 2025, we are operating at around about 1,000 wafer starts a year, every three months a new
13:24tape-out.
13:25At this point, 60,000 chips a year.
13:29Everything in Europe.
13:31We could do this anywhere else in the world, but I guess you get the point.
13:35And the great thing is that the algorithms that we are building do not require anything that other countries currently
13:42have,
13:43because it's a new paradigm.
13:44These non-near functions have largely been ignored in the past,
13:48because they were just incompatible with the CMOS stack.
13:52So with that, we not only get hands on the hardware,
13:56but we get this very valuable junction between software and hardware layer.
14:02That's photonic computing.
14:05If you want to work with us, we are installed.
14:08So photonic computing started to become popular last year.
14:13There are a lot of startups around the globe who start stepping into the same race,
14:18but we are the first to ship devices.
14:20So if you want to work with us, we are deployed in nearly all HPCs in Europe.
14:26And we are also, for the first time since the end of last year, deployed at one of the largest
14:32European cloud providers,
14:34IONOS, where you can book our service as a service.
14:39So it's no longer a paper discussion about sovereignty.
14:44It already happened.
14:46And you can start playing, working and operating with these systems.
14:50And with this, I'm at the end.
14:55The organizers said, if there are questions, there can be questions.
14:59If not, I hope you enjoyed what I've presented to you,
15:02and maybe you follow us somewhere and see the progress of Quant.
15:06Thank you very much.
15:15There is a question.
15:16What are your next steps?
15:17What are the next obstacles to take?
15:19Come to ISC, International Super...
15:22So there was a question, what are the next steps?
15:24Next week, we're demonstrating a really great step in terms of AI,
15:29because there is not just image recognition we run.
15:32I'm not...
15:33Wait till next week.
15:35There is a new, more advanced AI model being operated live on stage at the ISC.
15:39At the end of this year, we're shooting for LLMs.
15:42And going global.
15:53That's...
15:54Yeah.
15:58There is one more question.
16:03I don't know if you can explain it simply, but how does it work?
16:07You show the results, but how do you make it work?
16:13You mean...
16:17Okay, how does a processor work?
16:19I mean, this would be probably a one year's lecture, but the essence is you can think about
16:26light, and light is already a wave.
16:28Now, if you have a wave function, you can basically turn wave functions into wave functions
16:33at no cost.
16:35So when you, for instance...
16:37It's a bit of an analogon.
16:39Everyone wearing a glass or glasses here in the room
16:43has an energetic-free Fourier transformator on the nose.
16:47You can do the same on chip level.
16:50Full stop.
16:51You don't need energy.
16:52While a Fourier transformation in a digital world,
16:55it's millions of transistors equivalent to electricity, equivalent to data.
17:00It's energetic-free.
17:02That's one example.
17:04There are more functions like convolutions, like exponential functions,
17:08assigned functions.
17:09Everything is there if you understand how to basically control it on the chip level.
17:14This is what we do.
17:22Still two minutes.
17:26A small question about the competition.
17:31How do you position your competitors?
17:34I think I know Chinese also work on optronics and these fields.
17:41So how do you position yourselves in this?
17:46Is this exactly the same technology or different fundamentally?
17:52And your perspective?
17:54Okay.
17:56So, first of all, the pickup rate on this technology worldwide is great.
18:00We know the German competitors.
18:03There are some.
18:03There are some from Great Britain.
18:06They're in the US just recently.
18:08Two new startups have been founded.
18:09And the Chinese have ever since been strong on photonics.
18:12But it's great to have competition.
18:15So, I mean, that's the best thing that can happen, right?
18:18I mean, it motivates us highly.
18:20At this point, we can say from whatever public information we have,
18:27we're the only company who's able to ship full systems.
18:31So we have an advantage, temporal advantage.
18:34And that's where we want to lean in.
18:35So when you ask me what's the best protection against competitors,
18:38this is just move faster than everyone else.
18:41Because why?
18:43And this is an invitation to everyone who want to join on that race.
18:48The value is generated in the algorithm and the software part.
18:52We have a new tool.
18:53We build a new tool.
18:54We have a new chicken, basically.
18:57And you can now build new eggs.
18:59And the eggs are generating the value supported by the chicken.
19:03So that's the point.
19:04So the earlier you step into the race, the more likely it is that we can protect
19:08as a combination of algorithm and applications together with this technology.
19:17Up to the moderators.
19:18There are more questions.
19:26I would have actually two questions.
19:27But you can decide which one or whether you want to answer both.
19:31Since you're building analog computers,
19:34the first question would be how you tackle the kind of precision problem.
19:37And the second would be how universally applicable is your processor.
19:41Because usually you always have this trade-off between kind of efficiency
19:44for certain problems and how universally applicable it is.
19:47And that's a very gentle question.
19:49So the first one, as at least on reports,
19:52we are the only company reaching 16-bit floating point.
19:55In that respect, there is no issue on the accuracy level any longer,
20:00at least when it comes to AI.
20:01And on the second question, that's a good question.
20:13That's true.
20:16I don't know if I got it right how it works internally with all the wave function and all that.
20:23That doesn't mean that you can also solve some of the problems that quantum computing wants to solve.
20:29Because it's really complex, all the simulation of waves and all the interactions.
20:34So I'm saying something where probably most of the quantum scientists in the room are going to hate me for.
20:40Sorry, let's challenge on that.
20:43So the way I see it is quantum computers offers quantum effects.
20:47And it's great if your problem is a quantum problem.
20:50If your problem is not a quantum problem, let's go on a challenge.
20:54So what we cannot do, we can't basically exploit entanglement or superposition,
21:01or we are not working with Heisenberg's uncertainty relation and these kind of things.
21:05But that's the space where quantum computers are strong in, and this is where we should need them.
21:09And as long as your problem requires these functions, use a quantum computer.
21:13But if we look into classical hard math, I would take up the challenge here.
21:19Because you don't need to model things more complicated.
21:23So just on the mapping, you lose a lot of energy.
21:26And we can process things as you've ever written it in the formulas and math.
21:32That's the point.
21:33So this is where my current belief is.
21:36This is why I put up this ecosystem of a hybrid stack.
21:39Quantum computers are part, and they have to be there.
21:42But the question and the debate is still out.
21:44What has to be executed on a quantum computer?
21:47In my belief, it's especially where you can take large advantages of the quantum effects.
21:51If it's a complex math problem, let's fight.
22:00Right, I'm out of time.
22:04Thank you very much.
Comments