Photonic MAC Circuit Using Temporal Multiplexing for Large-K Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Photonic circuits for optical computing face limitations in performing large-scale vector-matrix multiplications due to the concurrent nature of operations, which restricts the size of the input vector (K) and increases power consumption, thereby limiting accuracy and efficiency.
Innovation Solution
Employing temporal multiplexing in a photonic circuit to decouple the number of source and modulator pixels from the size of the input vector (K), allowing for high accuracy modeling by using a single source pixel and N modulator pixels, where K is effectively unlimited, and achieving high power efficiency by operating at low power levels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If concurrent operations are used in photonic circuits, then computational performance is achieved, but the size of the input vector (K) is restricted and power consumption increases
Solution Approach 1:
The patent segments the concurrent operations into temporal sequences. Instead of requiring all K multiply operations to occur simultaneously using K source pixels and K modulator pixels, the system performs operations sequentially in time slots. Each time slot processes one multiply operation using a single source pixel and single modulator pixel, reducing the hardware footprint while maintaining computational performance through temporal multiplexing.
Solution Approach 2:
The patent implements periodic action by cycling through different source pixels and modulator pixels in a sequential manner. The controller systematically activates different pixel combinations across multiple time slots, creating a periodic pattern of operation that enables the system to process all K multiply operations using a limited set of pixels reused across time, thereby reducing the total number of pixels needed.
2Measurement precision
If larger K values are used to improve accuracy, then modeling accuracy increases, but power consumption increases
Solution Approach 1:
The patent segments the computational workload across time slots rather than requiring all operations to occur simultaneously. By dividing the K multiply operations into sequential time slots, the system can use a small number of low-power pixels for each operation while maintaining high K values for accuracy. The temporal multiplexing allows the system to achieve large K without proportionally increasing peak power consumption, as pixels are activated sequentially rather than all at once.
3Measurement precision
If more source pixels and modulator pixels are used to increase K, then accuracy improves, but device complexity increases
Solution Approach 1:
The patent implements universality by making each source pixel and modulator pixel multi-functional across different time slots. The same physical pixel can serve as a source pixel in one time slot and as a modulator pixel in another, or participate in different multiply operations at different times. This temporal multiplexing allows a small set of pixels to perform the work of what would traditionally require K source pixels and K modulator pixels, reducing device complexity while maintaining the ability to process large K values for high accuracy.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The solution enables high accuracy modeling and high power efficiency in performing multiply-and-accumulate operations for deep neural networks and other AI/ML algorithms by decoupling K from the number of pixels, allowing large K values and reducing power consumption.
Implementation Method 1
A source pixel generates a light beam with an intensity modulated according to values of the input row vector
Implementation Method 2
an optical fan-out structure configured to generate a plurality of copies of the light beam
Implementation Method 3
N modulator pixels respectively transmit the N copies to generate N transmitted light beams, wherein the N modulator pixels have individual transmissivities respectively modulated according to column vectors of the weight matrix
Implementation Method 4
N detector pixels respectively receive the N transmitted light beams and accumulate charge respectively in response to the N transmitted light beams
Data Source
AI summary
Various embodiments of the present disclosure are directed to a photonic circuit for a vector-matrix operation. A source pixel is configured to generate a light beam. An optical fan-out structure is configured to generate a plurality of copies of the light beam. A plurality of modulator pixels are configured to respectively transmit the plurality of copies with individual transmissivities to generate a plurality of transmitted light beams. A plurality of detector pixels are configured to accumulate charge respectively in response to the plurality of transmitted light beams. A controller is configured to control the source pixel and the plurality of modulator pixels to modulate an intensity of the light beam and the individual transmissivities to perform the vector-matrix multiplication operation. The intensity is modulated to temporally encode an input row vector, and the individual transmissivities are modulated to temporally encode corresponding column vectors of a weight matrix.