CIM Floating-Point MAC Circuit for Exponent-Difference Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing compute-in-memory (CIM) circuits for performing MAC operations on floating point numbers in machine learning suffer from accuracy issues due to the truncation of mantissa products with normal exponent differences when processing exponent differences of varying magnitudes, leading to compromised final sum accuracy.
Innovation Solution
A CIM circuit that separates the processing of mantissa products based on their exponent differences, handling those equal to or less than a difference threshold in parallel with those greater than the threshold, using dynamically configured thresholds to improve accuracy by preventing contamination of normal exponent differences by small exponent differences.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If mantissa products with varying exponent differences are processed together in a single pipeline, then processing speed is improved, but accuracy deteriorates due to truncation of normal exponent differences by small exponent differences
Solution Approach 1:
The patent divides the processing pipeline into two separate parallel pipelines: a first pipeline for handling mantissa products with exponent differences greater than a threshold, and a second pipeline for handling mantissa products with exponent differences less than or equal to the threshold. This segmentation prevents the truncation of normal exponent differences by small exponent differences, thereby maintaining accuracy while still achieving high processing speed through parallel operation of both pipelines.
2Device complexity
If a single processing pipeline is used for all mantissa products, then device complexity is reduced, but manufacturing precision deteriorates due to inability to handle different exponent distributions
Solution Approach 1:
The patent implements segmentation by creating two distinct processing pipelines with different configurations optimized for specific exponent difference ranges. The first pipeline is designed for normal exponent differences while the second pipeline handles small exponent differences, ensuring that each pipeline's processing characteristics match its intended input range, thereby achieving high computation precision without excessive complexity.
Solution Approach 2:
The patent introduces dynamic threshold configuration that allows the system to adapt to different distribution characteristics of exponent differences. The threshold between the two pipelines can be dynamically adjusted based on the input data characteristics, enabling the system to optimize its processing strategy for different computational scenarios while maintaining accurate results.
Data Source
AI summary
A computing-in-memory circuit includes an input circuit to receive a number (N) of input pairs, each of the N input pairs comprising a first one and a second one of N exponents, and a first one and a second one of N mantissas; a first adder circuit to generate N exponent sums based on the first and second exponents of the N input pairs; a subtractor circuit configured to calculate N exponent differences, each of the N exponent differences being equal to a difference between a corresponding one of the N exponent sums and a largest one of the N exponent sums; and a comparator circuit to compare each of the N exponent differences with a threshold to generate N control signals. N mantissa products of the first and second mantissas of the N input pairs, respectively, are to be selectively combined based on the N control signals.