Podcast EP334: The Unique Benefits of LightSolver’s Laser Processing Unit Technology with Dr. Chene Tradonsky

0:00Hello, my name is Daniel Nenny, founder of semi-wiki, the open form for semi-conductor professionals. Welcome to the semi-conductor insiders podcast series. My guest today is Shen Prodonsky, a physicist and the CTO and co-founder of Light Solver, where he leads the development of a proprietary physics-based computing system built on coupled laser dynamics to accelerate compute-heavy simulations and other computationally demanding workloads. Before moving into physics, he started in electrical engineering, a combination that helps him bridge advanced computing and complex physical systems. Welcome to the podcast, Shen. Thank you, Daniel, for having me. I'm really excited to be here and talk with you today. I'm excited as well. It's a great topic. So, start, please tell us a little bit about your background and your company, Light Solver.

1:00Well, I started in electrical engineering and then I moved into physics. While I'm working in my master degree at Technion, I worked on semi-conductor quantum dots, which give me strong foundation on quantum system and photonics. So, my interest positioned me at the intersection of engineering and complex physical systems. But when I think about what took me in that direction, well, I always been fascinating by this idea. Can we tackle computationally demanding problems like heavy simulation by taking physics-based approach, building the right physical system, system that mimicked the original process instead of forcing everything for digital computers. But while I was working in my PhD at Fights Manistitude,

2:01that all came together for me. I worked on a couple of days of network. Originally, I started my study on emerging network behavior like synchronization and pattern formation across different apologies. And those latent networks are actually controllable physical systems. And I realized that can mimic or simulate other physical system as an alternative approach to digital computing and much, much faster. I know I was behind academic curiosity. I was moving down the path to a new computing paradigm. And fortunately at Fights Man, I also met Wutim and Schlommi. I was working on real quantum system, specifically interactions of cold atoms and cold ions. And we share the same intuition that the physics-based dynamics could be the foundation of practical computational platform.

3:01And in 2020, we founded Light Software. And we built on the promise of developing these physics-based processors, built on capabilities dynamics. To accelerate the compute heavy simulations and other computation demanding workloads. That was the genesis of the laser processing unit or the LPU. Interesting. From what I understand, optical and photonic processors are being experimented with as accelerators for AI. But you've argued that one of its most important near-term uses will actually be in scientific simulation. What's drawing optical processors towards solving partial differential equations instead of just chasing AI workloads? That's an excellent question. Optics and photonic processor are absolutely being explored for AI. And for the core idea makes sense. Optics can do certain operations very efficiently.

4:03Like if we transform or vector matrix multiplication, the practical challenges is when large AI systems demands extreme scale, precision and stability. And in many photonics AI concept, you still need substantial electronic control calibration. And optoelectronics eye conversion. And when you look on the full system, those overheads can reduce or eliminate the end to end advantage. That's require careful co-design and still need some research breakthrough to reach a practicality and scale up. And scientific simulation is different. And this is where we focus. A large fraction of the HPC or high performance computing workloads are in partial differential equation based simulations. They sit at the core of engineering and science from, you know, fluid dynamics and heat transfer to a multiplicity systems like weather forecast and fusion reactors.

5:13And those types of simulations, our foundation of critical applications such as aircraft or automotive design and energy system and advanced manufacturing. And this is where we excel. Our laser processing unit or the LPU is not a photonic version of conventional computer. It's fully parallel physical system, all variables presents simultaneously. Inside is a couple days of cavity and the interactions are implemented as in Jimmy coupling between those days and light from one light laser to the other from probable fashion. So the state of the entire system evolved at once in hardware without constant memory fetch and data movement that you see in digital architecture. This is one of the major bottleneck in HPC that HPC faces today. In practice, the problem, the impact is you program the coupling of a specific PDE based task and the light physics evolved and thought for the stable state.

6:26And then you just read out the result. This improves both time to solution and energy efficiency for most demanding part of the task. Okay, so let's level set for listeners who don't live inside HPC every day. What are partial differential equations and why do they sit at the heart of so many simulations in industries like aerospace automotive and climate science. So partial differential equation or PDE is are the mathematical language that used to describe how physical quantities changes across space and time. Instead of single number changing with time that you describe an entire field like velocity pressure temperature or stress defined in defined everywhere in three dimensional space and evolving over time. That's why PDE sit at the heart of so many industrial scientific simulations. There flow over wind heat transfer in an engine mechanical stress in a structure, a climate and weather models and even parts of chip design like a heat dissipation and electromagnetic fields are already used to PDE with boundary condition and material properties.

7:46Good example is a turbulent airflow around the wind the wind is in the mid in meters in size by the thin boundary layer determine the dry and stall can be orders of magnitude smaller to capture both scales you discretize the space into very fine 3d mesh and step the solution forward in time. And each time step now require a solving large coupled equations that represent repeatedly until the solution is is stable. This is why PDE simulations consume enormous amount of energy and compute and why. As we increase the fidelity and resolution the computational cost goes dramatically often faster than the performance we gain from each new generation of hardware.

8:48So you mentioned that optical and photonic processors will start moving out of research labs, maybe this year and into real operational environments. What does that shift look like in reality and where do you expect these systems to show up first. So I don't think the moving to operational environment would look like you know sudden revolution or complete replacement of digital computing paradigm. It would look like a pragmatic insertion of existing into existing workflow where optical processors are used as specialized engines and are charged by measurable time to solution and energy to solution. I think we will continue to see a photonics adopted in IO and interconnect but the more interesting shift beyond IO is when photonics start contributing to computation itself. In practice, it would show what first as domain specific accelerator that plug into existing HPC pipeline rather than replacing them.

9:56Early deployment will be in small scale pilot and inside HPC centers with industrial simulation team focused on models that are most intensive in terms of front time and power consumption. And the key is not doing everything optically but offloading part of the parts where the physical system can evolve in parallel efficiency and deliver stable and repeatable outcome quickly. We also expect the physics based free space optical systems to appear early in those pilots because they can be packages appliances integrated with standard software workflow and evaluated in real industrial industrial problems with clear acceptance criteria. I think the winners would be the system that can integrate that the cleanly show a repeatable gains and come with realistic validation path.

11:05Okay, you also introduced the concept of physics native computing as a new category hardware. What does that term mean in simple terms and why is now the moment for it to emerge alongside CPU GPUs and even you know the becoming quantum systems. So let's explore that physics native computing many using controllable physical system in as a computing engine. Instead of simulating everything step by step on a digital processor you program the interactions of the physical substrate was dynamics matches the mathematics structure that you care about you let it evolve and then you read out the results. This concept is not new in spirit engineer have long use physical models as computers like a wind on us in aerodynamics or analog electronics circuits for dynamic dynamical system systems.

12:05What hold many of those approaches back from being accepted as general computing tools was programmability. If each new problem require building you physical setup you can scale beyond the narrow use case is changing now is significant we understand how building physical native systems can that are programmable and repeatable enough to be used across class of problems. This is the game changer here the the programmability optics is a good example in optics it enable you to represent many variables simultaneously and engineer the interactions between them and a reconfigurable way. So somehow the same hardware can can be adopted for different types of tasks rather than locked into a single task.

13:15I also now see this parallel to quantum computing one of the original motivation for quantum computing as a fine man noted back in the early 1980s was that simulate quantum systems that are extremely hard to compute classically. It's native computing follow similar high level principle but in our case the target is classical computing workloads. The point I want to emphasize is that there is very little overlap between those systems and you can think that you can choose to run on quantum computer and what we can run on our system. So both can imagine as a different complimentary approaches alongside CPUs and GPUs. So you're describing a gradual approach of integrating optical engine into existing HP infrastructure in which makes sense.

14:16How should simulation teams think about integrating optical hardware into their current tool chains to gain speed and energy efficiency without breaking what already works. That's that's a great question. There is no not overnight replacement you know of HPC the simulation team should not think about optical hardware as a new monolithic platform that forces rewrite everything. The winning pass is hybrid to introduce an optical engine as a complimentary compute component inside existing workflow simulation team should think in terms of partitioning in the computation. The digital side is still still does a lot of heavy lifting and matching discretizing discretization choices multi physics coupling parent post processing and verification and it also remain the system of record form regression test and accuracy matrices.

15:25The physical engine is then used selectively for the for the parts of the computation that matches the fence for example tightly coupled updates over many variables that can evolve simultaneously in physics native way. If iterative computation in your workloads for example Poisson or wave or Navier stocks equation if you were familiar with those types of partial differential equations. You can think of it as well defined compute primitive or a small repeatable computing task computing task inside the broader pipeline. You can measure end to end and time to solution and energy to solution while maintaining a clean fallback path to the digital baseline. The first that teams adopt incrementally they do not break what already works they add our accelerator the optical engine validate it on a substrate of workloads and expand usage only when it's stable scalable operationally simple enough to integrate.

16:43How do customers normally engage with you and your company I mean your website is light solver calm. Yes, what else can they do didn't touch with you so they can contact us through the website they can learn a lot of about us through the website there's a lot of information there and they can conduct us and we can give them access to additional material to learn how to work with the LPU. And they can use it as a background both for education and also contact us directly and walk on our LPU 100 we have LPU lab accessible already for earlier doctors so if it's very relevant to us also will give them access. And with that I know I want to first of all thank you for that for this opportunity. Yeah, thank you for being our guest it was a great conversation and I hope we can keep in touch let us know how your progress is going and know how to talk to you again soon.

17:48Yeah, it's really exciting to be here and thank you again. That concludes our podcast thank you all for listening and have a great day. Thank you.