Beyond Von Neumann Architectures: Exploring Algorithmic Opportunities via Octantis

Abstract

Today, one of the problems the scientific community is called upon to tackle is the well-known von Neumann bottleneck, which concerns the limitation in the bandwidth between the CPU and memory in a digital electronic system. Among the various solutions under study, the concept of Logic-in-Memory (LiM) has been proposed: a memory device that embeds simple computational elements between the different cells to define a distributed processing system. The present work introduces an extended version of Octantis, a novel open-source software useful for exploring LiM architectures. To achieve this goal, the internal structure of Octantis takes inspiration from the one of standard High-Level Synthesis tools, distinguishing itself from them for the target topology addressed. It analyses user-defined standard-C algorithms and determines which LiM architecture would be best suited to implement it. At its output, the tool provides a VHDL description of the synthesised circuit along with a custom test-bench. The earlier version of Octantis efficiently synthesised rather simple user-defined C algorithms. The version discussed here has been improved by extending the allowed complexity of input C-codes, like addressing nested loops and non-trivial data dependencies, and introducing hardware-specific optimisations to meet resource constraints. Several case studies have been considered to validate the newly implemented techniques and to analyse the capabilities of the tool in implementing data-intensive algorithms. The results demonstrate that Octantis can produce architectures that comply with the LiM topology while significantly reducing the exploration space to meet specific hardware requirements, such as memory dimensions and maximum logic integration. This methodology provides initial insights into potential LiM units that can be adopted in customised designs, making it a valuable tool in researching alternative electronic devices.