Flexible Multiplier for Neural Processing Element Precision Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural processing units face limitations in accurately handling high workloads for deep-learning tasks due to precision issues related to overflow and underflow, which affect the accuracy of data calculations.
Innovation Solution
A processing element with a flexible multiplier and saturating adder that can switch between different precision modes based on a mode signal to prevent overflow or underflow, allowing for dynamic precision conversion during calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data is processed with higher precision (more bits), then calculation accuracy is improved, but hardware resource consumption increases
Solution Approach 1:
The patent implements dynamic precision switching by providing multiple multiplier units with different precision capabilities (first precision and second precision) and selectively activating them based on overflow/underflow detection. This allows the system to adapt precision levels during operation rather than being fixed, resolving the contradiction between accuracy and hardware resources.
Solution Approach 2:
The system changes the precision parameter of data processing dynamically. When overflow or underflow is detected, the system switches from processing data in a first precision to processing in a second precision (or vice versa), thereby adjusting the calculation accuracy to match the actual data requirements and avoid unnecessary hardware resource consumption.
2Device complexity
If fixed precision is used for all calculations, then hardware implementation is simplified, but accuracy deteriorates due to overflow or underflow
Solution Approach 1:
The patent segments the multiplication function into multiple specialized multiplier units, each optimized for a specific precision level. Instead of using a single fixed-precision multiplier, the system divides the functionality into first multiplier units and second multiplier units, allowing accurate processing of both normal and edge-case data without compromising hardware simplicity.
Solution Approach 2:
The system achieves multi-functionality by incorporating multiplier units that can handle different precision requirements within the same hardware architecture. The first and second multiplier units work together to provide universal processing capability for various data ranges, preventing overflow and underflow while maintaining implementation feasibility.
3Reliability
If higher precision is always used, then overflow and underflow are prevented, but hardware resource efficiency decreases
Solution Approach 1:
The system dynamically adjusts precision usage based on actual data characteristics and overflow/underflow detection results. Rather than always using high precision, the system switches between first precision and second precision modes, ensuring reliability when needed while maximizing hardware resource efficiency during normal operation.
Solution Approach 2:
The precision parameter is changed adaptively based on detection results. When overflow or underflow is detected, the system switches to a different precision mode to prevent the error. This parameter change ensures reliability only when necessary, thereby maintaining hardware resource efficiency during normal processing.
Data Source
AI summary
A processing element, a neural processing device including the same, and a method for calculating thereof are provided. The processing element includes a weight register configured to receive and store weights, an input activation register configured to store input activations, a flexible multiplier configured to receive the weight and the input activation, to perform a multiplication calculation in a first precision or a second precision different from the first precision according to a mode signal, occurrence of an overflow, and occurrence of an underflow, and to generates result data; and a saturating adder configured to receive the result data and generate subtotals.