An optical chip that solves the Ising problem 100 times faster than a GPU Sally Ward-Foxton Optical transmission technology, the great "behind-the-scenes" of the information society.

 　Lightelligence, a U.S. start-up in optical computing, has demonstrated a silicon photonics accelerator. The company says it can solve Ising models more than 100 times faster than a typical GPU-based system.

　Lightelligence's photonic arithmetic computation engine, Pace, is an integrated optical computing system operating at 1 GHz, consisting of about 12,000 photonic devices. The system is said to be about a million times faster than the company's prototype Comet, which was announced in 2019 and consists of 100 photonic devices. Lightelligence also showed off a use case other than AI acceleration for its hardware for the first time in this latest demo.

Prototype of Pace [Click to enlarge] Source: Lightelligence

　Pace is said to run NP-complete problem-class algorithms, which are considered to be computationally extremely difficult, many times faster than existing accelerators. While it has not demonstrated optical superiority for all applications, it is capable of solving the Ising problem 100 times faster than a typical GPU. It is also 25 times faster than Toshiba's Simulated Bifurcation Machine (SBM), a dedicated system for Ising problems.

　The NP-complete problem has a very large state space and requires huge computing resources to solve it. This class also includes the Ising problem, the maximum cut/minimum cut problem, and the traveling salesman problem. Applications where NP-complete problems actually occur include bioinformatics, scheduling, circuit design, materials exploration, cryptography, and optimized power grids.

　We decided to demonstrate NP-complete acceleration because it exemplifies the advantages of optical computing," said Yichen Shen, CEO of Lightelligence, in an interview with EE Times.

　The core of our optical computing engine is that it can complete matrix multiplication in much less time than GPUs. 64×64 matrix multiplication takes less than 10 nanoseconds or about 5 nanoseconds, whereas GPUs take hundreds of clocks," Shen said. For NP-complete problems, where matrix multiplication is performed over and over again, the advantage of our technology is significant. We hope to use this new technology to find problems where photonics has an advantage," he claimed.

　The iterative nature of the NP-complete algorithm means that continuous matrix multiplication is dependent on prior results. This minimizes the bottlenecks that arise in the system electronics part. In other words, data does not have to travel back and forth through memory between multiplications.

　For larger commercial use cases, digital electronics and memory reads and writes can be a drag on the overall computing system," said Shen. We expect to achieve at least a few times faster, if not 100 times faster," he said.

Development of Photonics Technology

　Lightelligence is also working on developing photonics technology for data broadcasting and data interconnect to reduce bottlenecks.

External view of the Pace board, which is about the size of a PCIe card [Click to enlarge] Source: Lightelligence

　When asked if Lightelligence will aim to commercialize accelerators for NP-complete problems, Mr. Shen said, "As for hardware, we can try to enter the market. However, this technology will be applied to our products, so it will be able to address a wide range of markets, including AI (artificial intelligence) acceleration.

　Optical computing based on silicon photonics can increase computational speed and power efficiency by orders of magnitude. The technology is based on moving modulated infrared light into a silicon wire called a waveguide. The waveguide can be fabricated by applying a standard CMOS process. A kind of analog computing efficiently combines the two waveguides and the two signals while an on-chip modulator (which modulates the brightness) efficiently amplifies the two signals. At the same time, an optical MAC unit can be formed. However, while optical computing is great for accelerating linear operations such as matrix multiplication, it requires standard digital electronics for nonlinear operations, memory, and control.

　Lightelligence, like its competitor Lightmatter, uses a silicon photonics version of a Mach-Zehnder interferometer (MZI) as its arithmetic unit. However, while Lightmatter uses MEMS to change the physical shape of the MZI waveguide, Lightelligence injects electrons into the waveguide to adjust the optical refractive index and modulate the optical signal that passes through it.

　The Lightelligence technology, like other optical designs, has the potential to process multiple inputs simultaneously using different types of wavelengths and polarizations," Shen said. For example, different colors can be applied to a set of AI inferences," he said.

　The core chip of Pace, demonstrated by Lightelligence, is an ASIC-controlled die flip chip connected to a photonic die. This assembly is mounted via a PCB on an existing board and connected to the laser source by a fiber array. The mixed-signal ASIC has digital blocks in the control logic to control data flow and I/O as well as SRAM for data storage, etc. The analog part of the ASIC bridges the digital blocks and photonic devices.

　Maurice Steinman, vice president of engineering at Lightelligence, said, "Designing individual chips is very difficult, but integrating them is even more difficult. Because optical computing is really a form of analog computing, it requires a huge amount of circuit design, simulation, iteration, and test chips to achieve high fidelity results.

(Translation: Rumi Tanaka; Editing: EE Times Japan)

[Related links

Optical fiber to support digital innovation

Optical Computing Processors to Accelerate Inference

Two Companies Collaborate to Develop Devices for Next-Generation High-Capacity Optical Communications