Software-Defined Number Formats: Bridging the Gap between Performance, Accuracy, and Security

Información del proyecto

SoftNum

Identificador del acuerdo de subvención: 101031148

DOI

10.3030/101031148

Proyecto cerrado

Fecha de la firma de la CE 15 Marzo 2021

Fecha de inicio 1 Septiembre 2022

Fecha de finalización 16 Enero 2025

Financiado con arreglo a

• EXCELLENT SCIENCE - Marie Skłodowska-Curie Actions

Coste total

€ 224 933,76

Aportación de la UE

€ 224 933,76

Coordinado por

THE QUEEN'S UNIVERSITY OF BELFAST
United Kingdom

Periodic Reporting for period 1 - SoftNum (Software-Defined Number Formats: Bridging the Gap between Performance, Accuracy, and Security)

Período documentado: 2022-09-01 hasta 2024-08-31

Resumen del contexto y de los objetivos generales del proyecto

Nearly all computing devices implement the same standardized hardware-supported number formats to aid software portability. The MSCA SOFTNUM project attempts to answer the question of whether using the same standardized number formats for all applications is still the best approach to achieve high performance. The efficiency benefits of low-precision number formats suggest that a single format may no longer be sufficient for all diverse computing applications. In fact, a single application might benefit from using multiple software-implemented compact number formats, making the use of a single hardwired format no longer effective.
The main objective of the SOFTNUM is to explore software-defined number formats, an innovative approach which transforms number formats, generally considered to be hard-wired functionality, software-defined. These software-defined number formats are designed at the software level to meet the numeric needs of the computation and are more compact representations of the data. They reduce memory footprint and bandwidth requirements without sacrificing accuracy, which results in a speedup compared to the purely use of conventional hardwired number formats. This new approach can have significant positive outcomes for society. Reducing computations and memory storage using flexible software-level small numbers leads to sustainable computing. This means a smaller carbon footprint and lower energy requirements while still providing better speed and the same accuracy as traditional methods.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

The fellow has successfully designed the software-defined number formats and developed a comprehensive theoretical and experimental study of feasibility and effectiveness of this innovative way of designing number formats. The SOFTNUM has showed that for an application like Belief Propagation algorithm, which is used in fields like machine learning, communications, and robotics, traditional hardware-supported floating-point formats (either 64, 32 or even 16 bits) can be replaced with a customized 8-bit floating-point format at the software level, resulting in over a 70% speed improvement compared to the 64-bit floating-point format. This significant achievement has been accepted for publication in IEEE Transactions on Emerging Topics in Computing, a reputable venue for publishing computing research (https://ieeexplore.ieee.org/document/10847799).
Beyond the Belief Propagation algorithm, the fellow identified and removed numerical redundancy in the CKKS Encoding part of Homomorphic Encryption and published the results at a reputable annual security symposium, 19th ACM Asia Conference on Computer and Communications Security (https://dl.acm.org/doi/10.1145/3634737.3656292). Moreover, the fellow participated in a collaborative survey paper, contributing a comprehensive review of existing number formats and their role in the resource efficiency of convolutional networks. This contribution was published in Section 4 of a survey paper in the well-known ACM Computing Surveys (https://dl.acm.org/doi/pdf/10.1145/3587095).
Moreover, the fellow disseminated the SOFTNUM project results by presenting a poster covering achievements in both machine learning and security applications at an important annual computing workshop at Imperial College London (the poster titled "Software-Defined Number Formats: Bridging the Gap between Accuracy, Performance and Security" was presented at the Novel Architecture and Novel Design Automation Workshop (NANDA): https://www.doc.ic.ac.uk/~phjk/NANDA24/). Furthermore, the fellow has communicated the research achievements with the general public during the Computer Science open days outreach events at Queen’s University Belfast. He engaged with potential candidates and their families, explaining how research benefits society and emphasizing the significance and impact of studying and researching computer science.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

The SOFTNUM project has introduced a novel paradigm in which number formats are software-defined rather than hardwired. This advancement goes beyond traditional computing, which relies on conventional hardware-supported floating-point number formats. In such systems, concerns about portability and compatibility can lead to unnecessary computational and memory overhead for many applications.
This project has demonstrated that by leveraging software-defined number formats, computational efficiency can be achieved without sacrificing accuracy, as shown by comprehensive theoretical and experimental validations published in the resulting papers. By reducing numerical format size and optimizing numerical data communications, the project contributes to reducing the carbon footprint of large-scale computations, aligning with global sustainability goals. Furthermore, the project's contributions to Homomorphic Encryption will have an impact on confidential AI and efficient secure computing for health and financial data processing.

The proposed software-defined floating-point (FP) number formats and their performance comparison