Optical Chip is Faster than GPU

Although a typical GPU setup can solve the Ising problem with ease, now a silicon photonics accelerator can also do the same, but at a speed a hundred times faster. The optical computing startup Lightelligence has demonstrated this feat.

The photonic arithmetic computing engine from Lightelligence is an integrated optical computing system, and it is known as Pace. It consists of about twelve thousand photonic devices that run at 1 GHz each. Compared to Comet, the earlier 100-device prototype from Lightelligence that they unveiled in 2019, Pace has a speed advantage of 1 million times. This is the first time that Lightelligence has demonstrated a use case on its hardware that goes beyond AI acceleration.

Lightelligence has designed Pace to run algorithms for problems that belong to the NP-Complete class. These represent one of the most difficult computational issues, requiring much higher speed systems compared to existing accelerators. Pace did not demonstrate optical superiority for all applications. However, it beat a typical GPU when executing the Ising problem by a factor of 100. In fact, it was even defeated by a factor of 25 a system that Toshiba assembled especially for solving the Ising problem—the simulated bifurcation machine running on FPGAs.

With a huge state space, NP-complete problems require very large computing resources for tackling them. The computing time depends on a polynomial of the size of the problem, scaling in proportion. This class includes the Ising problem, traveling salesman problem, and the graph max-cut/min-cut problem. In reality, NP-Complete problems can be found in scheduling, bio-informatics, material discovery, circuit design, power grid optimization, and cryptography applications.

According to their CEO Yichen Shen, Lightelligence decided to demonstrate the acceleration of NP-Complete problems as this best illustrated the advantages of optical computing.

The chief advantage of the optical compute engine from Lightelligence is it can compute matrix multiplications much faster than GPUs can. Typically, GPUs take several hundreds of clock cycles to complete a 64 x 64 matrix multiplication. According to Lightelligence, Pace can do it in about 5 nsec. As NP-Complete problems require several iterative matrix multiplications, Pace has the upper hand. Lightelligence wanted a problem that best demonstrated the superiority of this new technology.

The major factor for Pace is the iterative nature of the algorithms that the NP-Complete problems use. Moreover, the successive matrix multiplications depend on the result of the previous calculations. In GPUs, system electronic parts cause the bottleneck, as data must shuttle to and from the memory in between multiplications. With bigger commercial use cases, the read and write cycles in digital electronics increases tremendously such that the entire computing system slows down. Lightelligence is confident it will be able to demonstrate advantages at least several times faster, if not 100 times.

Optical computing has numerous advantages. Based on silicon photonics, Its main advantage is its speed—several orders of magnitude improvements in power efficiency and computing speed. Basically, the system directs modulated infrared light within silicon wires or waveguides. Scientists accomplished this by using standard CMOS processes.