Photonic Tensor Core Matrix Vector Multiplier

High-computation speed is a need of an hour in today’s artificial intelligence era. In existing general-purpose processors, matrix operations occur serially, while requiring continuous access to the cache memory, thus creating the “von Neumann bottleneck” problem. The revolutionizing impact of Neural Networks (NNs) have contributed to development of emerging technologies, ranging from free space diffractive optics to Nano photonic processors that aim to improve the computational efficiency of specific tasks performed by NNs.

Researchers at GW have invented a Photonic Tensor Core (PTC) processor unit. The system consists of a photonic dot product engine (PDPE) that receives input(s) and is configured to compute optical and/or electro-optical tensor operations of the input(s) by performing optical or electro-optical, or all-optical dot-product multiplications, thus performing multiply-accumulate (MAC) operations. The entire PTC processor is comprised of modular PTC sub-modules, which perform the MAC operations. The PTC sub-modules consists of a PDPE having first inputs(s) and second input(s), which are matrix, vector or a scalar. The PTC and PDPE have integrated photonics, fiber optics, optical free-space, and/or combination of these.

Figure 1. Tensor core unit conceptual processor used to multiply and accumulate 4x4 (Convolutional Neural Network)

Applications:

• Enable real-time intelligent computing at the edge of ultra-high-speed mobile networks (5G and beyond) and internet-connected devices.
• Can be employed as micro data centers (DC) in edge devices.
• Intelligent Data Traffic Routing (e.g. for Data Centers).
• Intelligent Pre-filtering & Correlations in Very Large Array (VLA) Systems.
• Network Security: anomaly & misuse detection.

Advantages:

• High speed computing
• Low latency
• Highly efficient matrix multiplication with 8-bit precision