Optical computing startup Lightelligence has developed a new Photonic Arithmetic Computing Engine (PACE) that outperforms NVIDIA‘s most powerful Turing-based GPU in the GeForce RTX 3080 by nearly 100x in NP-complete problems.
Incorporating 3D packaging and seamless co-design efforts into its technology, Lightelligence remains the only company with fully integrated optical computing systems working at speed. Under the Ising model, Lightelligence’s new PACE accelerator is demonstrating record-breaking results in understanding phase transitions.
Sign up for your weekly dose of what's up in emerging technology.
Inside Lightelligence’s PACE
Lightelligence’s Photonic Arithmetic Computing Engine (PACE) uses standard silicon photonics integration of a Mach-Zehnder Interferometer (MZI) for computation and MEMS (micro-electromechanical system) to change the shape of the waveguide in the MZI. By injecting electrons into the waveguide, PACE modulates its photonic refractive index, which modulates the optical signal that passes through it.
An ASIC control die flip-chip is bonded to a photonic die at the centre of the chip. A fibre array assembly is used to connect to the laser source, and this assembly is mounted on a conventional substrate using a PCB. The mixed-signal ASIC includes a digital block with control logic, I/O, and SRAM for storing data. Digital and photonic devices are connected by the analogue part of the ASIC.
Lightelligence uses 12,000 optical devices integrated into a circuit clocked at 1GHz. According to Shen, the founder and CEO of Lightelligence Inc, Lightmatter’s technology can simultaneously process a variety of inputs by using different wavelengths or polarisations of light (such as using different colours for a pair of AI inferences at the same time).
Potential applications of Lightelligence’s PACE
This new accelerator can solve complicated mathematical problems at 25 to 100 times the speed of conventional computers. In addition, it can also be used in material science, thermodynamics, bioinformatics, cryptography, circuit design, power grid optimisation, and much more.
Solving the latency problem of AI hardware
PACE supports the PCI-e interface and is designed to work with current hardware and AI software frameworks such as Google TensorFlow and Facebook PyTorch. When AI models are deployed in the physical world, low latency can be a game-changer, especially in the field of neural network computations. This new accelerator improves latency, throughput and power efficMatrix multiplication. It is also able to speed up neural network computations and thus help create faster AI models.
The optical computer has many advantages, including high density, small size, low junction heating, high speed, dynamic scaling and reconfigurability into smaller and larger networks/topologies, and massive parallel computing. Lightelligence’s optical processor makes a significant contribution to the bright feature of optical computing.