In Ansys’ case, the engineering and physics simulation startup used LightSolver’s LPU simulation to accelerate graph partitioning in its LS-Dyna physics simulation package.
Sign up now to receive highlights, analyses and stories from the week straight to your inbox without any interference.
As the whole world is still wrapped up in AI, there are loads of high-performance computing tasks waiting to be tackled, and a speedup in any one area could make a significant difference, whether it’s computational fluid dynamics, material analysis, or something else.
To tackle substantial problems, LightSolver envisions deploying entire clusters of LPUs where work can be distributed crosswise.
Presently, LightSolver’s prototype comprises 100 lasers. Nonetheless, by 2027, the Israeli startup anticipates having an LPU with 200 lasers, and by 2029, scaling it up to 1,000 lasers.
While the advantage of more effective graph partitioning may be somewhat limited in traditional physics simulations, Banerjee foresees significant benefits from technologies like the LPU in workloads such as cell placement, commonly seen in electronic design automation (EDA) software. As you might recall, Ansys is currently in the process of being acquired by Synopsys, a major supplier of such software.
Nonetheless, optimization problems are just one of the various applications that LightSolver is presently exploring.
Per Banerjee, in 80 percent of instances, LightSolver’s LPU simulation achieved a better solution, which made simulations about 15 to 20 percent faster.
Of course, this is assuming someone successfully assembles enough qubits to create a quantum computer capable of practical use.
This whole process takes place within microseconds. According to Lightsolver, for optimization problems, the limiting factor is the number of variables each LPU can address being proportional to the size of the laser network.
“Compared to this, the LPU performs computations using lasers with massive parallelism, allowing all potential options to be checked simultaneously,” as explained by Lightsolver.
It may not seem like much, but “if LS-Dyna could be accelerated by 20 percent, that would be a significant improvement,” he noted. He also highlighted that, at least for physics simulations, quantum computing still holds the potential for much greater speedups, possibly up to a million times faster.
The laser light emitted from the network enters a cavity with two mirrors facing each other. Inside, the beams interfere with one another as they pass through a gain medium. This medium amplifies the laser light so that it can be reverted back to a mathematical solution by a detector camera.
We’ve noticed quantum computing vendors, such as Intel, Microsoft, and Google, adopting a comparable strategy as they strive to establish systems with a sufficient number of logical qubits for practical use.
But as the challenges in engineering get tougher and the margins slimmer, companies like Ansys are always looking into new technologies, like generative AI, machine learning, and even quantum processing units, to step up the pace.
These problems are so complex that it’s generally quicker to approximate a solution within acceptable limits using heuristics like trial and error, rather than solving the equation.
For now, LightSolver has crafted a software framework to simulate its LPU on GPUs. This simulation offers potential customers a platform to assess larger problems than the current hardware could manage.
- After several iterations
- back and forth between the mirrors
- the top solution is the one that receives the greatest amplification, the company clarified.
The startup is also looking into how its LPU could expedite solving partial differential equations (PDEs), a common aspect of computational fluid dynamics and finite element analysis.
“Right now, we’re really pushing forward in the world of HPC, using shared memory, message passing, and GPUs,” shared Prith Banerjee, the chief technology officer of Ansys, in a recent chat with The Next Platform.
These technologies won’t be taking over traditional computing; rather, they’ll be complementing them, similar to how GPUs have been used to ease specific tasks off the CPUs. One of the latest technologies that Ansys is experimenting with is a laser-based computing platform from an Israeli startup called LightSolver.
At its core, LightSolver’s LPU comprises a network of lasers where problems are encoded by adjusting each laser’s phase and amplitude.
If you aren’t familiar, graph partitioning is employed in large-scale simulations to split up the workload into smaller segments for parallel processing. Ideally, the graph is divided in a way that minimizes the edges that cross the cut. This cut is known as the min-cut. The closer you can get to this cut, the more efficient the workload parallelization would be, resulting in faster performance.
The CEO of LightSolver, Ruti Ben-Shlomi, likens each laser to a stone being thrown into a pond. The waves interfere with one another, eventually forming a single wave that represents the final solution.
- Optimization problems like these have traditionally been a focus for fringe computing platforms such as quantum processing units, analog computing, and optical computing, and LightSolver’s LPU is no different. Unlike quantum systems that often need power-hungry cryogenic facilities to diminish noise interference with the qubits, LightSolver’s LPU runs at room temperature, is entirely analog, and works on traditional principles
- no dive into the quantum world required.
- Ansys is specifically investigating if LightSolver’s laser processing unit (LPU) can speed up solving non-polynomial equations
- often known as NP-hard
- typically found in physics simulations. The compute requirements for these problems grow exponentially as the number of variables increases, making it impractical to find a definitive solution.
For these tasks, the scaling is somewhat different, with the laser network capable of handling significantly more grid points. By 2027, Lightsolver aims to have an LPU capable of solving PDEs with 100,000 grid points, and a million grid points by 2029.
In a typical heuristic search, the optimization problem is broken down into a series of yes/no questions. The more variables there are, the more operations are needed to explore all possible combinations for the best solution.
