Switched Charge Matrix for Programmable Neural Crossbar Connectivity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current analog implementations of machine learning face challenges such as difficulty in producing programmable switch matrix connectivity, bandwidth limitations, high power consumption, and the need for significant tuning and tweaking.
Innovation Solution
The use of switched charge circuits, which allow for rapid reconnection of communications paths and do not rely on physical connections or device values, enabling efficient and dynamic neural network operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If switched capacitor circuits are used for analog machine learning implementation, then capacitor matching and finite bandwidth devices are required, but this leads to difficulty in producing programmable switch matrix connectivity and significant tuning requirements
Solution Approach 1:
The patent replaces the mechanical/physical capacitor-based switching system with a photonic system using light pulses to control charge transfer. Instead of using physical capacitor connections and switches that require precise matching, the invention uses optical signals to gate charge movement between nodes, eliminating the need for capacitor matching and reducing tuning requirements.
Solution Approach 2:
The invention changes the fundamental operating parameter from voltage/capacitance relationships to charge quantity relationships. By using charge as the primary information carrier and controlling it via photonic gating, the system achieves programmable connectivity without the precision requirements of capacitor matching, as charge quantity can be precisely controlled through current integration over time.
2Measurement precision
If translinear loops are used for analog multiplication, then significant setup time is required, but this wastes power and limits bandwidth
Solution Approach 1:
The patent employs periodic photonic gating signals to control charge transfer in discrete time intervals. Instead of requiring continuous setup time as in translinear loops, the system uses periodic light pulses to gate charge movement, enabling rapid switching between operations and improving bandwidth while maintaining multiplication accuracy through precise temporal control.
Solution Approach 2:
The invention maintains continuous charge storage nodes that can rapidly respond to incoming charge packets without requiring setup time. The photonic gating mechanism allows immediate control of charge transfer when activated, eliminating the setup time waste of translinear loops and enabling continuous high-speed operation with improved power efficiency.
3Ease of operation
If fixed impedance multiplexers and crossbars are used for switch matrix connectivity, then device impedance interferes with device modulation results, but this causes distortion requiring tuning measures
Solution Approach 1:
The patent replaces the electrical impedance-based switching mechanism with photonic gating. Instead of using fixed impedance multiplexers that interfere with signal modulation, the invention uses light-controlled charge transfer gates that do not introduce impedance-related distortion, thereby improving modulation control and eliminating the need for tuning measures.
Solution Approach 2:
The invention introduces photonic signals as an intermediary control mechanism between the switch matrix control logic and the charge transfer process. This photonic intermediary does not interfere with the charge signals being modulated, unlike fixed impedance multiplexers, and provides clean, distortion-free control of the switching operation.
4Adaptability or versatility
If triode mode transistors are used for weight modulation, then high power consumption occurs even for short periods, but this reduces energy efficiency
Solution Approach 1:
The patent uses periodic photonic gating to control weight modulation, where transistors are activated only during brief light pulse intervals rather than remaining in high-power triode mode. This periodic activation achieves the necessary weight programming flexibility while dramatically reducing average power consumption compared to continuous triode operation.
Solution Approach 2:
The invention maintains weights in a low-power state between operations and only activates power-consuming elements during the brief periods when weight modulation is actually needed, controlled by photonic signals. This eliminates the waste of keeping transistors in high-power triode mode even for short periods, while preserving full weight programming flexibility.
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
This approach allows for high-speed data processing, reduced power consumption, and improved accuracy by eliminating the need for capacitor matching and finite bandwidth devices, while also enabling efficient calibration and operation of neural networks.
Implementation Method 1
a pinned photodiode to produce an input charge
Implementation Method 2
a transfer gate to transfer the input charge to a second charge storage device
Implementation Method 3
At least one first current source is coupled to the at least one output charge storage device. At least one second current source is coupled to the shared node connecting the at least one input charge storage device and the at least one output charge storage device proportional in magnitude to the at least one first current source
Data Source
AI summary
A reconfigurable, for example with time, network switch matrix coupling switch charge circuits representing multiply and add circuits (MACs) and neurons (MACs with activations) capable of accepting and outputting proportional to charge pulses through crossbars within said network, said crossbars controlled by local controllers and higher level controllers to setup said crossbar communications.