18674632. ASYNCHRONOUS ACCUMULATOR USING LOGARITHMIC-BASED ARITHMETIC simplified abstract (NVIDIA Corporation)

ASYNCHRONOUS ACCUMULATOR USING LOGARITHMIC-BASED ARITHMETIC

Organization Name

NVIDIA Corporation

Inventor(s)

William James Dally of Incline Village NV (US)

Rangharajan Venkatesan of San Jose CA (US)

Brucek Kurdo Khailany of Austin TX (US)

Stephen G. Tell of Chapel Hill NC (US)

ASYNCHRONOUS ACCUMULATOR USING LOGARITHMIC-BASED ARITHMETIC - A simplified explanation of the abstract

This abstract first appeared for US patent application 18674632 titled 'ASYNCHRONOUS ACCUMULATOR USING LOGARITHMIC-BASED ARITHMETIC

The abstract describes a method for adding logarithmic format values efficiently in neural networks, which typically involve convolution layers that require multiplication and addition operations.

• Logarithmic format values are added by decomposing exponents into quotient and remainder components, sorting based on remainders, summing sorted quotients using an asynchronous accumulator, and multiplying partial sums by remainders to produce a sum.
• This approach simplifies addition of logarithmic format values compared to conventional methods that involve converting values to integers and back to logarithmic format.
• The method aims to improve energy efficiency in neural networks by optimizing the addition process for logarithmic format values.

Potential Applications: - This innovation can be applied in various deep learning models that utilize convolution layers, such as image recognition, natural language processing, and speech recognition. - It can also be beneficial in edge computing devices where energy efficiency is crucial for prolonged operation.

Problems Solved: - Simplifies the addition of logarithmic format values in neural networks, reducing computational complexity and improving energy efficiency. - Streamlines the process of handling logarithmic format values in convolution layers, enhancing overall performance of deep learning models.

Benefits: - Energy-efficient addition of logarithmic format values in neural networks. - Improved computational efficiency in convolution layers, leading to faster processing speeds. - Enhanced accuracy and performance of deep learning models utilizing logarithmic format values.

Commercial Applications: - Optimizing neural network operations in edge computing devices for improved energy efficiency. - Enhancing the performance of deep learning applications in various industries such as healthcare, finance, and automotive.

Questions about Logarithmic Format Value Addition: 1. How does the method of adding logarithmic format values in neural networks compare to traditional methods in terms of efficiency? 2. What are the potential implications of using this approach for energy-efficient computing in edge devices?

Frequently Updated Research: - Stay updated on advancements in logarithmic format value addition techniques in neural networks to ensure optimal performance in deep learning applications.

Original Abstract Submitted

Neural networks, in many cases, include convolution layers that are configured to perform many convolution operations that require multiplication and addition operations. Compared with performing multiplication on integer, fixed-point, or floating-point format values, performing multiplication on logarithmic format values is straightforward and energy efficient as the exponents are simply added. However, performing addition on logarithmic format values is more complex. Conventionally, addition is performed by converting the logarithmic format values to integers, computing the sum, and then converting the sum back into the logarithmic format. Instead, logarithmic format values may be added by decomposing the exponents into separate quotient and remainder components, sorting the quotient components based on the remainder components, summing the sorted quotient components using an asynchronous accumulator to produce partial sums, and multiplying the partial sums by the remainder components to produce a sum. The sum may then be converted back into the logarithmic format.