Multi-Injection Phase Summing Circuit for PVT-Robust Neural Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural network implementations face challenges in reducing power consumption and surface area while being sensitive to PVT variations and requiring complex auxiliary circuits for frequency domain reading, especially in remote server-based calculations.
Innovation Solution
A multi-injection phase summing circuit using an oscillator locked by injection interface circuits with synchronized signals out of phase, encoding information in the phase domain to reduce power consumption and surface area, and making measurements less sensitive to PVT variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If analog and mixed blocks (DAC converter and comparator) are used to produce weighted sum, then power consumption is reduced and area is reduced, but sensitivity to PVT variations increases
Solution Approach 1:
The patent replaces traditional analog DAC converters and comparators with a phase-based oscillatory system. Instead of using analog voltage/current comparisons that are sensitive to PVT variations, the invention uses phase information from oscillators to represent and process neural network weights and inputs. This substitution of the computational mechanism eliminates the sensitivity to process, voltage, and temperature variations while maintaining low power consumption and small area.
Solution Approach 2:
The invention changes the domain of computation from analog voltage/current to phase domain. By encoding neural network parameters (weights, inputs, outputs) as phase differences between oscillators rather than analog magnitudes, the system achieves immunity to PVT variations. The phase information remains stable despite changes in process, voltage, or temperature, solving the reliability issue while keeping power consumption low.
2Reliability
If phase-locked loop circuits are used per neuron, then robustness to PVT variations is improved, but surface area is penalized due to passive loop filter
Solution Approach 1:
The patent extracts and removes the bulky passive loop filter component from the phase-locked loop circuit. Instead of using a complete PLL with its large filter, the invention uses a simplified oscillatory neuron model where phase information is naturally maintained through the oscillator's operation. This extraction of the problematic component achieves PVT robustness without the surface area penalty of traditional PLL filters.
Solution Approach 2:
The invention uses multiple identical oscillators whose phases are coupled through injection locking rather than using complex PLL circuits for each neuron. By copying the simple oscillatory behavior across multiple neurons and using phase coupling to achieve synchronization and computation, the system achieves PVT robustness with minimal area, avoiding the need for large loop filters in each neuron.
3Productivity
If nanomagnetic oscillators are used, then integration density is improved and energy efficiency is improved, but reading output signals becomes difficult and requires auxiliary latch detection circuitry
Solution Approach 1:
The patent merges the computation and reading functions into a unified phase-based system. Instead of using nanomagnetic oscillators that produce small voltage signals requiring separate detection circuitry, the invention uses oscillators where the phase information itself is the computational result. The phase differences between oscillators directly represent the neural network outputs, eliminating the need for auxiliary latch detection circuitry while maintaining high integration density.
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 achieves reduced power consumption and surface area while making information more easily exploitable and less sensitive to PVT variations, enabling efficient weighted sum calculations in neural networks.
Implementation Method 1
an oscillator locked by a plurality of injection interface circuits (501, 502, 503) controlled by synchronization signals at the same frequency but out of phase with each other
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
Multi-injection phase summing circuit comprising an oscillator (300) having a natural oscillation frequency F0, a reference injection interface circuit (701) controlled by a reference synchronization signal at a reference frequency and at least one additional injection interface circuit (702,703) controlled by a secondary synchronization signal at the reference frequency, each injection interface circuit (701,702,703) having a variable injection parameter, the secondary synchronization signals each being phase-shifted with respect to the reference synchronization signal, the oscillator (300) being configured to generate an output signal at the reference frequency and phase-shifted with respect to the reference synchronization signal by a phase shift that is a function of a sum of the respective phase shifts of each secondary synchronization signal with respect to the reference synchronization signal, weighted by the injection parameters.