Floating-Point CIM Post-Multiplication Alignment Through Exponent Phasing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computing-in-memory (CIM) circuits face challenges in accurately performing MAC operations on floating point numbers due to potential truncation of mantissa products when shifting to a maximum exponent sum, leading to reduced accuracy and increased power consumption.
Innovation Solution
A CIM circuit that divides floating point numbers into multiple phases based on the exponent sum range, aligning and accumulating mantissa products within each phase to enhance accuracy and reduce power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating point numbers are shifted to maximum exponent sum for alignment, then alignment accuracy is improved, but mantissa product truncation occurs leading to reduced computational accuracy
Solution Approach 1:
The patent segments the floating point numbers into multiple phases based on their exponent sums. Instead of shifting all numbers to the maximum exponent sum simultaneously (which causes truncation), the system divides the alignment process into multiple phases where numbers are shifted to different target exponent sums. This segmentation allows each mantissa product to be aligned without excessive truncation, resolving the contradiction between alignment accuracy and computational accuracy.
2Device complexity
If traditional single-phase alignment is used, then device complexity is low, but power consumption increases due to unnecessary computations
Solution Approach 1:
The patent introduces a dynamic phase selection mechanism that adapts the alignment process based on the distribution of exponent sums in the input data. The system determines the number of phases and their specific target exponent sums dynamically during runtime, rather than using a fixed single-phase approach. This dynamic adaptation allows the system to perform only the necessary alignment operations, reducing power consumption while maintaining manageable complexity through automated phase determination.
3Ease of operation
If mantissa products are shifted to maximum exponent sum, then alignment is simplified, but truncation of less significant bits occurs increasing power consumption
Solution Approach 1:
The patent applies local quality by assigning different target exponent sums to different phases based on the local characteristics of the input data distribution. Instead of uniformly shifting all numbers to the maximum exponent sum, the system identifies local clusters of exponent sums and creates phases tailored to each cluster's characteristics. This allows each phase to perform alignment with minimal truncation, reducing the need for power-intensive correction operations while maintaining alignment simplicity within each phase.
Data Source
AI summary
A computing-in-memory circuit (CIM) circuit includes an input circuit to receive N first inputs and N second inputs; N summing circuits, each configured to combine the corresponding first exponent and the corresponding second exponent of a corresponding one of the N input pairs to generate a corresponding one of N exponent sums; a first selector circuit configured to select a largest exponent sum; a phasing circuit configured to divide at least a portion of the N exponent sums into N phases from the largest exponent sum, each of the N phases associated with a respective one of N exponent subsets; and N subtractor circuits, each configured to calculate a corresponding one of N exponent differences for each of the N phases, each of the N exponent differences equal to a difference between a corresponding one of the N exponent sums from the respective exponent subset and the largest exponent sum.