Architecture Specialization from arIthmetical and MemOry perspectiVes (ASIMOV)

Principal Investigators: Alberto A. del Barrio and Jos├ę Luis Ima├▒a.
Start and end dates: 01/09/2022-31/08/2025
Code: PID2021-123041OB-I00

Summary: Nowadays, Computer Science is witnessing how the post-Moore Era is getting close to its extinction as major vendors are approaching the physical limits of CMOS technology. In addition, carbon footprint restrictions imposed by the European Union and others make it difficult for the Computer Architecture community to keep the path of the evermore demanding software applications, so it is time for different approaches.

This technological slowdown has been countered in recent years via heterogeneity and specialization. In the ASIMOV project we leverage the specialization hypothesis in a three-fold manner to contribute with efficient solutions under stringent power budgets. First, given a concrete kernel and application domain, we propose to evaluate arithmetic and memory subsystem improvements to optimize the features of the kernel. Second, we pursue the development of hw/sw platforms where these solutions can be integrated in order to provide both usability and programmability, as well as giving a wholesome measurement of the proposed optimizations. Third, unconventional approaches have been appearing in the last years and, although technologically they may not be mature enough yet, it is important to evaluate them in concrete domains and compare with classical solutions to study the potential gains in the near future.

At the arithmetic level, the ASIMOV project is focused on two types of computations: the Next Generation Arithmetic (NGA) and the cryptography arithmetic. The NGA encompasses new non-standard formats in combination with approximate computing techniques, with special emphasis in Artificial Intelligence (AI) applications. The cryptography arithmetic is the basis for two cutting-edge security use cases, namely: Blockchain at the hardware level in IoT scenarios, and post-quantum cryptography to face the challenges arisen from the disruptive appearance of Quantum Computing.

From the memory perspective, ASIMOV proposes a holistic approach. First, the interface with the datapath leverages light data types and structures, such as the NGA types, to reduce the memory footprint. Second, within the memory subsystem itself, new Non-Volatile Memory (NVM) technologies, such as STT-MRAM or SoT-MRAM, are the basis for the presented hierarchies. At the organizational level a combined hw/sw cache management is proposed in combination with tuned replacement policies and prefetching strategies.

In order to provide programmability and usability to the developed specialized components, we intend integrating them into a hw/sw platform and release some of these under public licenses. Standalone accelerators, RISC-V-based cores and SoCs through ISA extensions and simulators will be our main target platforms to perform such integration. Finally, from a higher level of abstraction, the ultimate goal of the project is to incorporate all these macro-components into a framework to automate the deployment. On the one hand, this will enable the exploration of multiple solutions and, on the other hand, it will speed up the process.

Finally, as an alternative to traditional computation, ASIMOV envisages what we call unconventional approaches with great potential in the upcoming years. Under this umbrella, the project pays special attention to analog and quantum computing.

Objectives:
Objective-1 (O1). Specialized Arithmetic and Approximate Computing
O1.1. Next Generation Arithmetic and Approximate Computing
O1.2. Cryptographic Arithmetic
Objective-2 (O2). Domain specific memory system specialization
Objective-3 (O3). Integration of Specialized Components
Objective-4 (O4). Exploration of Unconventional Architectures