Inner Product Device Using Index Unit and Storage Operation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional devices for computing inner products in neural networks require significant computational resources and memory, leading to high power consumption and inefficient usage of hardware architecture due to unbalanced computation and memory access.
Innovation Solution
A device for computing inner products that incorporates an index unit, a storage operation unit with computing-in-memory architecture, a redundant to 2's complement converter, a mapping table, and a multiplier-accumulate module, which uses index values to reduce computation amount and leverage parallel accumulation capabilities for balanced hardware usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional MAC-based inner product computation is used, then computation can be performed, but computation amount is very large and power consumption is high
Solution Approach 1:
The patent replaces conventional electronic multiplication and accumulation operations with a physical field-based computation system using voltage signals and capacitive coupling. The inner product computation is achieved through parallel voltage signal propagation and natural capacitive charging/discharging processes, eliminating the need for traditional MAC units and significantly reducing computational complexity and power consumption.
Solution Approach 2:
The patent segments the weight data into multiple groups corresponding to different capacitive coupling paths. Each group of weights is processed simultaneously through dedicated voltage signal paths, enabling parallel computation of multiple inner product terms. This segmentation transforms sequential MAC operations into concurrent physical processes.
2Productivity
If conventional MAC-based computation is used, then inner product can be computed, but computation amount is very large and memory requirements increase
Solution Approach 1:
The patent transitions from digital domain computation to an analog voltage domain, adding a physical dimension to the computation process. Weight values are represented as voltage levels rather than digital numbers, and inner product computation is performed through parallel voltage signal interactions governed by capacitive coupling, fundamentally changing the computational paradigm and reducing memory burden.
Solution Approach 2:
The capacitive coupling structure performs computation automatically through its inherent physical properties. When voltage signals are applied, the capacitors naturally charge and discharge according to the weighted connections, performing multiplication and accumulation operations without requiring external computational control. The physical structure itself executes the computation.
3Device complexity
If representative weight accumulation is used, then computation amount is reduced, but hardware usage is unbalanced and efficiency is low
Solution Approach 1:
The patent assigns different capacitive coupling strengths to different weight groups, creating local variations in coupling characteristics. Each group of capacitors has optimized coupling parameters tailored to its specific weight range and computational role. This local optimization ensures that each part of the hardware contributes effectively to the overall computation, achieving balanced and efficient resource utilization.
Data Source
AI summary
A device for computing an inner product includes an index unit, a storage operation unit, a redundant to 2's complement (RTC) converter, a mapping table, and a multiplier-accumulate (MAC) module. The index unit, storing index values, is coupled to word lines. The storage operation unit includes the word lines and bit lines and stores data values. The mapping table stores coefficients corresponding to the index values. The index unit enables the word line according to a count value and the index value, such that the storage operation unit accumulates the data values corresponding to the bit lines and the enabled word line, thereby generating accumulation results. The RTC converter converts the accumulation results into a total data value in 2's complement format. The MAC module operates based on the total data value and the coefficient to generate an inner product value.