Enhanced Block Floating Point Multiplier for Neural Network Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Block Floating-Point (BFP) representation is inadequate for applications like Convolutional Neural Networks (CNNs) due to its limited range, leading to inaccurate results when dealing with widely differing numeric distributions, as small numbers are often set to zero, limiting the representation of large data ranges.
Innovation Solution
The Enhanced Block Floating Point (EBFP) format stores exponent differences and significands more efficiently, allowing for the representation of small numbers and wider ranges by using a variable allocation of payload bits between exponent and significand, enabling accurate representation of numbers that would otherwise be zero in BFP.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional Block Floating-Point representation is used to reduce storage resources and computing complexity, then device complexity and storage requirements are reduced, but the range of representable numbers is limited and small numbers are lost (set to zero)
Solution Approach 1:
The patent applies dynamics by making the number format configurable between different precision modes (e.g., 8-bit, 16-bit, 32-bit) depending on the computational requirements. The system dynamically adjusts the representation format based on whether accuracy or resource efficiency is prioritized, allowing small numbers to be preserved when higher precision modes are selected.
Solution Approach 2:
The patent changes the parameter of numerical representation precision by introducing configurable bit allocations for sign, exponent, and significand fields. This allows the system to adjust the range and precision of representable numbers, enabling small numbers to be accurately represented when the precision parameters are set appropriately.
2Reliability
If higher precision number representations are used to improve computation accuracy, then accuracy is improved, but storage resources are exceeded
Solution Approach 1:
The system dynamically selects the appropriate precision level (8-bit, 16-bit, 32-bit) based on the specific computational task and available storage resources. This allows the application to achieve necessary accuracy without consistently using the highest precision format, thereby optimizing storage utilization.
Solution Approach 2:
The patent enables parameter adjustment of the number format's precision characteristics, allowing the system to configure the balance between storage efficiency and computational accuracy. By changing parameters such as exponent and significand bit allocations, the system can adapt to different storage constraints while maintaining adequate precision.
3Device complexity
If BFP format with fixed precision is used to simplify the number representation, then device complexity is reduced, but the ability to represent large data ranges is limited
Solution Approach 1:
The patent creates a universal number representation system that can function in multiple precision modes (8-bit, 16-bit, 32-bit) within a single framework. This multi-functional design allows the same hardware or software infrastructure to handle both simple and complex numerical ranges, enhancing adaptability without proportionally increasing complexity.
Solution Approach 2:
The system dynamically adapts its representation capabilities by switching between different precision configurations. When large data ranges are required, the system activates higher precision modes with larger exponent fields, while maintaining simpler modes for less demanding applications, thus achieving versatility without permanent complexity overhead.
Data Source
AI summary
A data processing apparatus is configured to determine a product of two operands stored in an Extended Block Floating-Point format. The operands are decoded, based on their tags and payloads, to generate exponent differences and at least the fractional parts of significands. The significands are multiplied to generate an output significand and shared exponents and exponent differences of the operands are combined to generate an output exponent. Signs of the operands may also be combined to provide an output sign. The apparatus may be combined with an accumulator having one or more lanes to provide an apparatus for determining dot products.