Recently, the Logic-in-Memory (LiM) concept has been widely studied in the literature. This paradigm represents one of the most efficient ways to solve the limitations of a Von Neumann’s architecture: by placing simple logic circuits inside or near a memory element, it is possible to obtain a local computation without the need to fetch data from the main memory. Although this concept introduces a lot of advantages from a theoretical point of view, its implementation could introduce an increasing complexity overhead of the memory itself, leading to a more sophisticated design flow. As a case study, Binary Neural Networks (BNNs) have been chosen. BNNs binarize both weights and inputs, transforming multiply-and-accumulate into a simpler bitwise logical operation while maintaining high accuracy, making them well-suited for a LiM implementation. In this paper, we present two circuits implementing a BNN model in CMOS technology. The first one, called Out-Of-Memory (OOM) architecture, is implemented following a standard Von Neumann structure. The same architecture was redesigned to adapt the critical part of the algorithm for a modified memory, which is also capable of executing logic calculations. By comparing both OOM and LiM architectures we aim to evaluate if Logic-in-Memory paradigm is worth it. The results highlight that LiM architectures have a clear advantage over Von Neumann architectures, allowing a reduction in energy consumption while increasing the overall speed of the circuit.
13 November 2021 in IEEE Transactions on Computers
The increasing complexity of real-life applications demands constant improvements of microprocessor systems. One of the most frequently adopted microprocessor design scheme is the von Neumann architecture. Central Processing Unit (CPU performs computations and communicates with memory in a constant exchange of information. This unceasing motion of data between these two components became a significant performance bottleneck. A lot of power, energy, and computational time are wasted in this communication. With Beyond von Neumann Computing (BvNC paradigms, calculations are performed inside or very close to a memory array. BvNC approaches are proposed in the literature, mainly based on modifications of existing memories, enabling simple computations. Others exploit emerging technologies to both store and compute data, using analog operations. In this work we follow a different approach, where computational units are placed close to memory cells, improving versatility and performance. We propose a Hybrid-SIMD architecture made of memory and computing elements in an interleaved structure. Hybrid-SIMD can be used both as a low density memory and as SIMD accelerator. We insert our design in a classical von Neumann system based on a RISC-V processor, and we estimate its impact, demonstrating its capability to improve speed reducing at the same time energy consumption.
Most modern CPU architectures are based on the von Neumann principle, where memory and processing units are separate entities. Although processing unit performance has improved over the years, memory capacity has not followed the same trend, creating a performance gap between them. This problem is known as the "memory wall" and severely limits the performance of a microprocessor. One of the most promising solutions is the "logic-in-memory" approach. It consists of merging memory and logic units, enabling data to be processed directly inside the memory itself. Here we propose an RISC-V framework that supports logic-in-memory operations. We substitute data memory with a circuit capable of storing data and of performing in-memory computation. The framework is based on a standard memory interface, so different logic-in-memory architectures can be inserted inside the microprocessor, based both on CMOS and emerging technologies. The main advantage of this framework is the possibility of comparing the performance of different logic-in-memory solutions on code execution. We demonstrate the effectiveness of the framework using a CMOS volatile memory and a memory based on a new emerging technology, racetrack logic. The results demonstrate an improvement in algorithm execution speed and a reduction in energy consumption.
Nowadays, the modern electronic systems are facing an important limitation in terms of performance, known as von Neumann bottleneck. It affects the communications between two crucial elements, the CPU and the memory, which suffer from a saturation in bandwidth. Many solutions are currently under investigation and among them the concept of Logic-in-Memory (LiM) has been introduced: a memory enriched in its array of computational elements which enable the implementation of a flexible distributed processing system. The current work introduces Octantis, a High-Level Synthesizer useful for the exploration of LiM architectures. The proposed software analyzes an input algorithm described in standard C language and identifies which LiM architecture would implement it better. At its output, the synthesized solution is provided together with a test-bench, to properly characterize it, in terms of performance, spatial occupation and power consumption. Many algorithms have been successfully synthesized by Octantis and some of the results achieved will be discussed along the document.
In recent years, Computer-Aided Design (CAD) software have become indispensable tools for designing, testing, and synthesizing logic circuits. Commercial software provided by companies like Synopsys, Cadence, or Mentor dominate the Electronic Design Automation (EDA) field. Nevertheless, several Open Source tools are also available, and some of them aim at assisting the designer in emerging Beyond-von-Neumann computing paradigms, such as Logic-in-Memory (LiM). LiM is a promising architectural and technological solution to the von Neumann Bottleneck, i.e. the performance gap between the CPU and the memory in classical CPU-Memory structures. In literature, various approaches to the LiM paradigm have been proposed. This paper introduces the Design Explorer for In-Memory Architectures (DExIMA) tool, which has the ambitious aim of providing a well-defined design flow strategy for the development, validation and performance estimation of a wide range of LiM architectures. Currently, DExIMA focuses on Coarse-grain Logic-in-Memory (CGLiM) architectures, which integrate memory and computation elements at a coarse-grain level. Nevertheless, DExIMA encompasses a flexible architectural model and a modular performance estimation engine that can be adapted to LiM implementations where memory and logic elements are more finely integrated. Hence, DExIMA is a versatile tool offering an environment for testing and comparing different LiM solutions, empowering designers to explore novel approaches in LiM architecture design.
The amazing development of transistor technology has been the main driving force behind modern electronics. Over time, this process has slowed down introducing performance bottlenecks in data-intensive applications. A main cause is the classical von Neumann architecture, which entails constant data exchanges between processing units and data memory, wasting time and power. As a possible alternative, the Beyond von Neumann approach is now rapidly spreading. Although architectures following this paradigm vary a lot in layout and functioning, they all share the same principle: bringing computing elements as near as possible to memory while inserting customized processing elements, able to elaborate more data. Thus, power and time are saved through parallel execution and usage of processing components with local memory elements, optimized for running data-intensive algorithms. Here, a new memory-mapped co-processor (MeMPA) is presented to boost systems performance. MeMPA relies on a programmable matrix of fully interconnected processing blocks, each provided with memory elements, following the Multiple-Single Instruction Multiple Data model. Specifically, MeMPA can perform up to three different instructions, each on different data blocks, concurrently. Hence, MeMPA efficiently processes data-crunching algorithms, achieving energy and time savings up to 81.2% and 68.9%, respectively, compared with a RISC-V-based system.