Computing-in-Memory Device Signed Input Conversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional computing-in-memory (CIM) devices are limited to performing operations on single-bit or positive multibit inputs, failing to handle signed multibit inputs required by advanced neural network activation functions like leaky ReLU and Swish, which restricts their application in complex tasks.
Innovation Solution
A CIM device with an input conversion module that inverts the most significant bit of signed multibit input data to convert it into unsigned data, allowing MAC operations within the CIM, and an output conversion module that compensates for computational errors to produce equivalent results as if signed inputs were used, thereby enabling efficient neural network computations on signed multibit inputs without increasing device size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CIM devices perform operations on signed multibit inputs, then adaptability to advanced neural network activation functions is improved, but device complexity increases
Solution Approach 1:
The patent introduces an input conversion module as an intermediary that transforms signed multibit input data into unsigned multibit data before processing. This mediator allows the existing unsigned-only CIM device to handle signed inputs by converting them to a compatible format, performing MAC operations, and then converting the results back to signed format with appropriate compensation.
Solution Approach 2:
The patent changes the parameter representation of input data from signed to unsigned format during processing. By inverting the most significant bit of signed input data, the system transforms the data into a format suitable for existing unsigned CIM operations, performs computations, and then applies compensation to restore the correct signed result.
2Adaptability or versatility
If CIM devices perform operations on signed multibit inputs, then functionality for complex neural network tasks is improved, but structural changes increase device size
Solution Approach 1:
The patent segments the processing function into distinct modules: an input conversion module that handles signed-to-unsigned transformation, the existing unsigned-only CIM core for MAC operations, and an output conversion module that applies compensation. This segmentation allows the core CIM device to remain unchanged while adding functionality through modular converters.
Solution Approach 2:
The patent makes the CIM device universal by enabling it to process both unsigned and signed multibit inputs through conversion modules. The input conversion module accepts signed data and transforms it into unsigned format that the existing CIM can process, while the output conversion module restores the signed format, allowing the device to serve multiple input types without requiring separate hardware for each.
3Adaptability or versatility
If conversion modules are added to handle signed inputs, then compatibility with signed activation functions is improved, but computational complexity increases
Solution Approach 1:
The patent applies preliminary action by performing bit inversion on the most significant bit of signed input data before the MAC operation. This pre-processing step converts the signed input into an equivalent unsigned representation that can be processed by the existing CIM architecture, avoiding the need for complex signed arithmetic during the core computation.
Solution Approach 2:
The patent converts the limitation of unsigned-only processing into a benefit by using the input conversion module to transform signed inputs into unsigned format. The 'harm' of being unable to process signed data directly is converted into a benefit through systematic conversion and compensation, allowing the device to handle signed activation functions like leaky ReLU and Swish while maintaining efficient unsigned MAC operations.
Data Source
AI summary
Provided are a computing-in-memory (CIM) device and a neural network device with a computational layer. The CIM device includes an input conversion module configured to receive signed input data and convert the signed input data into unsigned input data, a CIM including multiple memory cells for separately storing weights and configured to receive the unsigned input data, perform a multiply-accumulate (MAC) operation between the unsigned input data and the stored weights, and output output data, and an output conversion module configured to output compensated output data by compensating the output data for a computational error. Accordingly, it is possible to efficiently perform neural network computation on a signed multibit input while minimizing an increase in size, computational complexity, and structural changes.