Floating-Point Format Conversion for Precision-Power Tradeoffs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
General-purpose processors are limited to computing floating-point numbers of a single format, leading to unnecessary accuracy and increased power consumption, especially in applications like neural networks where varying levels of accuracy are required.
Innovation Solution
A floating-point number processor that converts floating-point numbers between different bit lengths, allowing for flexible computation by adjusting the bit lengths of the exponent and mantissa fields based on calculated exponent bit lengths and base values, enabling efficient processing with reduced power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point numbers of high accuracy (greater bit length) are used, then measurement precision is improved, but use of energy increases
Solution Approach 1:
The patent applies dynamics by making the bit length of floating-point numbers adjustable rather than fixed. The processor can dynamically change the exponent bit length and mantissa bit length based on the computational requirements of different applications. This allows the system to use higher precision (more bits) when needed for accuracy-critical operations and lower precision (fewer bits) when sufficient for the task, thereby optimizing the trade-off between measurement precision and energy consumption.
Solution Approach 2:
The patent implements parameter changes by allowing modification of the floating-point number format parameters, specifically the exponent bit length and mantissa bit length. The processor can switch between different configurations (e.g., IEEE 754 single precision with 8-bit exponent and 23-bit mantissa, or half precision with 5-bit exponent and 10-bit mantissa) to match the requirements of different computational tasks, thus adjusting both precision and energy usage accordingly.
2Measurement precision
If floating-point numbers of high accuracy (greater bit length) are used, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent applies universality by designing a floating-point number processor that can handle multiple formats and precision levels within a single device. The processor is configured to support different exponent bit lengths and mantissa bit lengths, allowing it to perform both high-precision and low-precision computations. This multi-functionality eliminates the need for separate processors for different precision requirements, thereby managing device complexity while providing flexibility in accuracy.
3Device complexity
If floating-point numbers of a single format are used, then device complexity is reduced, but adaptability worsens
Solution Approach 1:
The patent applies dynamics by enabling the processor to adapt its configuration dynamically based on the computational workload. The system can switch between different floating-point formats (e.g., changing exponent and mantissa bit lengths) to match the requirements of different applications such as neural network training versus inference, thus achieving high adaptability without permanently increasing device complexity.
Data Source
AI summary
Aspects for converting floating-point numbers in a processor are described herein. As an example, the aspects may include receiving, by a floating-point number converter, an exponent bit length, a base value, and one or more first floating-point numbers of a first bit length. Further, the aspects may include calculating, by the floating-point number converter, one or more second floating-point numbers of a second bit length based on the exponent bit length and the base value, the one or more second floating-point numbers respectively corresponding to the one or more first floating-point numbers.