Computação Versátil em Processadores com Arquiteturas Avançadas

The project aims at exploring, developing, and demonstrating how specialized hardware on modern devices, and even specialized accelerators with a focus on AI, can be efficiently repurposed in the context of heterogeneous systems for general-purpose computing tasks. The main objectives of the project are the following: 1) Creation of a Modular Open Source Framework: We plan to provide a comprehensive means to programmers to speed up the development of applications using state-of-the-art features and systems. This is to be materialized as a modular open source framework with high interoperability with different devices and vendors. The framework will help programmers in exploiting both fine-grained and coarse-grained parallelism, hardware-specific features, as well as incorporate adaptive scheduling approaches targeting large-scale systems (including EuroHPC supercomputers). 2) Showcase Applicability Across Various Use Cases: One of the key goals is to showcase applicability, benefits, versatility and impact of repurposing AI-optimized hardware across different kernels and programs for broad computational needs. As use cases, we will rely on two bioinformatics applications (variant calling and epistasis detection) and on sparse matrix processing computations performed by scientific computing, data analytics, and simulation workloads from different domains (e.g. physics / chemistry modeling, recommender systems, risk analysis). 3) Exploration of Next-Generation Hardware Accelerators: Another goal of the project is to explore next-gen hardware accelerators to achieve unprecedented performance for the use case applications. In that line of research, we are going to focus on two breakthrough technologies: enhancing an existing RISC-V pipeline with specialized instructions to further boost performance and energy-efficiency, and exploring ReRAMs for workloads that are memory-bound, which is often the case in sparse computing. 4) Energy Efficiency and Sustainability: Because AI-specialized devices are present to a high degree, it becomes very important to take into account energy-efficiency. Thus, we plan to evaluate this metric and the sustainability implications of utilizing some of the more specialized hardware features / devices for general-purpose computing, contributing to more environmentally sustainable computing practices. 5) Development of software tools and libraries: The project execution will result in the development of tools related to the framework, the project use cases, and the exploration of next-gen hardware. The tools developed during the project will be made publicly available to researchers at a public code repository (e.g. GitHub), as well as any programmers with interest in using specialized hardware for general-purpose tasks. 6) Dissemination of research findings and tools: Finally, we will share our findings, and methodologies with the research community and industry through publications, open-source contributions, and/or other dissemination activities (e.g. workshops / tutorials). By achieving these objectives, the project sets itself to contribute to a paradigm shift in how modern hardware is perceived and used, expanding its applicability beyond AI to a wide range of computational challenges, enhancing computational efficiency and accessibility, and stimulating innovation across various fields. Thus, contributing to the advancement of scientific research, technology development, and societal progress.