Floating-Point CIM Exponent Mantissa Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Computing-in-Memory (CIM) processors are limited in performing floating-point computations due to processing delays and high energy consumption, as they are primarily designed for fixed-point operations, making them unsuitable for high-speed and energy-efficient deep neural network training.
Innovation Solution
A floating-point computation apparatus and method that separates exponent and mantissa computations, where the exponent processing is performed by a CIM processor and the mantissa processing is done by a dedicated mantissa processing unit, reducing processing delays and energy consumption by minimizing memory access and optimizing bitline precharging, and skipping unnecessary normalization steps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If a conventional CIM processor is used for floating-point computation, then memory access power consumption is reduced, but processing speed decreases significantly (5000 times slower than floating-point system)
Solution Approach 1:
The patent divides floating-point computation into two separate parts: exponent computation and mantissa computation. The exponent processing unit handles exponent addition and comparison operations using CIM, while the mantissa processing unit handles multiplication and normalization operations using high-speed digital logic. This segmentation allows each unit to be optimized for its specific function, achieving both energy efficiency and high processing speed.
2Use of energy by moving object
If floating-point computation is implemented in a CIM processor, then energy efficiency improves, but device complexity increases due to heterogeneous operations
Solution Approach 1:
The patent separates the complex floating-point computation into two distinct processing units with specialized functions. The exponent processing unit deals with simple addition and comparison operations, while the mantissa processing unit handles multiplication and normalization. This segmentation reduces device complexity by assigning specific tasks to dedicated units rather than requiring a general-purpose floating-point unit.
Solution Approach 2:
The patent introduces a data preprocessing unit that acts as an intermediary between the input data and the exponent/mantissa processing units. This unit separates the input floating-point numbers into their exponent and mantissa components, preparing them for parallel processing. The intermediary unit simplifies the overall system architecture by handling data preparation and coordination.
3Ease of manufacture
If a homogenous floating-point CIM architecture is used, then ease of manufacture improves, but processing speed decreases due to simple CIM processor limitations
Solution Approach 1:
The patent creates a hybrid architecture that segments the computation into exponent processing (handled by simple CIM) and mantissa processing (handled by high-speed digital logic). This segmentation allows the use of simple, easily manufactured CIM processors while achieving high processing speeds through the dedicated mantissa processing unit.
Solution Approach 2:
The patent applies different processing qualities to different parts of the floating-point computation. The exponent processing uses energy-efficient CIM with simple logic, while the mantissa processing uses high-speed digital logic with complex operations. Each part is optimized with the appropriate level of complexity for its specific computational requirements.
Data Source
AI summary
Disclosed herein are a floating-point computation apparatus and method using Computing-in-Memory (CIM). The floating-point computation apparatus performs a multiply-and-accumulation operation on pieces of input neuron data represented in a floating-point format, and includes a data preprocessing unit configured to separate and extract an exponent and a mantissa from each of the pieces of input neuron data, an exponent processing unit configured to perform CIM on input neuron exponents, which are exponents separated and extracted from the input neuron data, and a mantissa processing unit configured to perform a high-speed computation on input neuron mantissas, separated and extracted from the input neuron data, wherein the exponent processing unit determines a mantissa shift size for a mantissa computation and transfers the mantissa shift size to the mantissa processing unit, and the mantissa processing unit normalizes a result of the mantissa computation and transfers a normalization value to the exponent processing unit.