Differential Analog MAC Using Charge Transfer to Cut Power
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multiplier-accumulator architectures for machine learning applications face challenges in scalability and power consumption due to synchronous operation and increased gate complexity, particularly in performinng multiply-accumulate operations, which leads to high power dissipation and inefficiency.
Innovation Solution
An asynchronous multiplier-accumulator architecture is developed, utilizing a unit element with AND-groups and charge transfer capacitors to minimize internal state changes and power consumption, allowing for cascaded operations and efficient conversion of analog multiplication products to digital outputs through a charge summing unit and analog-to-digital converter.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If synchronous clocked stages are used for multiplication operations, then operational timing and synchronization are improved, but power dissipation increases due to continuous clocking
Solution Approach 1:
The patent employs periodic charge transfer cycles where capacitors are sequentially switched to transfer charges representing multiplication results. This periodic action replaces continuous synchronous clocking, achieving necessary timing control while reducing power dissipation by activating circuits only when needed for charge transfer operations.
Solution Approach 2:
The patent substitutes the mechanical synchronous clocking system with an electrical charge-based asynchronous system. Instead of using clock signals to coordinate operations, the system uses charge transfer through capacitors and voltage level transitions to signal completion of multiplication, eliminating the need for continuous clock-driven operation and reducing power consumption.
2Productivity
If large-scale multiplier-accumulator arrays are implemented for machine learning, then computational capability is improved, but gate complexity increases proportionally to n2
Solution Approach 1:
The patent uses charge copies instead of direct signal connections. Each multiplication result is represented as a charge quantity copied onto capacitors, which can then be transferred and accumulated without requiring complex interconnections between gates. This charge copying mechanism enables scalable multiplier-accumulator arrays where computational capability increases linearly with the number of units rather than quadratically.
Solution Approach 2:
The patent introduces capacitors as intermediary elements between multiplication operations and accumulation. These capacitors temporarily store charge representations of multiplication results and facilitate their transfer to summing nodes. This intermediary charge-based communication simplifies the interconnection architecture, reducing gate complexity while maintaining high computational capability through cascaded multiplier-accumulator stages.
3Use of energy by moving object
If internal state changes are minimized for static weighting matrices, then power consumption is reduced, but operational flexibility decreases
Solution Approach 1:
The patent implements dynamic reconfigurability through control signals that can enable or disable specific AND gate groups and adjust capacitor connections. For static weighting matrices in machine learning, the system minimizes internal state changes by maintaining stable capacitor connections, reducing power consumption. When adaptability is needed, control signals can dynamically reconfigure the circuit to accommodate different weighting matrices or computational requirements.
Solution Approach 2:
The patent applies different operational modes to different parts of the circuit based on local requirements. For static weighting matrix operations, the majority of the circuit operates in a low-power stable state with minimal switching. Control logic locally manages which AND gate groups are active and which capacitor connections are engaged, allowing the system to minimize overall power consumption while retaining the capability to activate specific regions when operational flexibility is required.
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 architecture reduces power consumption and enables scalable, low-power multiply-accumulate operations by minimizing internal state changes and using asynchronous operation, effectively addressing the inefficiencies of prior art multipliers in machine learning applications.
Implementation Method 1
each AND gate output coupled to an analog charge line associated with the bit order through a charge transfer capacitor of value Cu
Implementation Method 2
the charge summing unit having a first terminal of a respective charge summing capacitor coupled to a respective analog charge line according to the bit weight of the analog charge line
Data Source
AI summary
A differential multiplier-accumulator accepts A and B digital inputs and generates a dot product P by applying the bits of the A input and the bits of the B inputs to respective positive and negative unit elements comprised of groups of AND gates coupled to charge transfer lines through a capacitor Cu. Each positive and negative unit element receives one bit of the B input applied to all of the AND gates of the unit element, and each positive and negative unit element having the bits of A applied to each associated AND gate input of each unit element. The AND gates are coupled to charge transfer lines through a capacitor Cu, and the charge transfer lines couple to binary weighted charge summing capacitors and to an analog to digital converter to generate a digital output product. The charge transfer lines may span multiple unit elements.