DNN Bit-Inversion Compression for Faster Sparse Multiplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As deep neural networks (DNNs) become larger and more complex, the amount of data movement and calculation increases, leading to decreased calculation speed and increased energy consumption in memory and control devices.
Innovation Solution
A calculation device that inverts and represents some bits based on the sign of data values, compresses data by converting negative values to 2's complement, and performs multiplication calculations using recovery data values to reduce data movement and calculation time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If deep neural networks become larger and more complex to improve performance, then learning and prediction capability is improved, but calculation speed decreases and energy consumption increases
Solution Approach 1:
The patent extracts and processes only the significant bits (non-zero bits) of data values, separating them from the zero bits. By identifying and operating only on the extracted non-zero portions of data, the system reduces the amount of data that needs to be moved and processed, thereby improving calculation speed while maintaining the learning and prediction capability of large neural networks.
2Reliability
If deep neural networks become larger and more complex to improve performance, then learning and prediction capability is improved, but energy consumption increases
Solution Approach 1:
The patent extracts only the non-zero bits from data values and processes only these extracted portions. This extraction approach reduces the total volume of data that needs to be moved between memory and processing units, directly reducing energy consumption while preserving the essential information needed for accurate learning and prediction in large neural networks.
3Productivity
If sparse acceleration technology is used to reduce calculation amount, then calculation efficiency is improved, but data representation complexity increases
Solution Approach 1:
Instead of representing sparse data by storing only the positions and values of non-zero elements (traditional sparse format), the patent inverts the approach by inverting the bits of data values and processing only the inverted non-zero bits. This inversion simplifies the data representation by converting sparse data into a format where zeros become ones and vice versa, allowing for more efficient processing while reducing representation complexity.
Data Source
AI summary
Disclosed is the calculation device of a deep neural network, which includes a conversion unit that outputs a plurality of conversion data values including bits corresponding to bits included in a plurality of original data values and further including a sign bit based on a sign of each of the plurality of original data values each composed of a plurality of bits, a compression unit that generates compression data, and a calculation preparation unit that outputs a plurality of recovery data values by inverting a value of at least some bits in each of the plurality of conversion data values, and wherein the conversion unit determines an MSB of each of the plurality of original data values as the sign bit, and inverts values of bits included in each of original data values whose signs are negative among the plurality of original data values.