UGA IRGA - 2021-2022
UGA IRGA - Projet exploratoire et émergent
Society development has yielded new application domains (such as IoT, AI, image processing, Big Data) which require extremely large computing efficiency at low resources (e.g. energy). Further development of these applications is stagnated due to the limitations of current computing architectures. As a result, the computing paradigm has shifted towards dedicated accelerators in heterogeneous architectures. However, they have also hit a roadblock as the gain obtained from accelerators and direct implementation of the algorithm/computation in hardware using existing Von-Neumann architectures implemented in volatile technologies may stale soon. This is where the full potential of computing in-memory (CIM) using emerging technologies can be explored.

There is a very wide variety of CIM solutions proposed today that exploit existing technologies. They enable logic and/or arithmetic operations directly inside the memory boundaries. The operations are performed without the need of transferring data to/from the CPU, thus saving time and energy. This can be achieved exploiting the physical characteristics of the memory and/or inserting computational elements in the peripheral logic (sense amplifiers). However, due to ever-increasing number of resistive-memory devices and applications being developed very limited research has been carried out to analyze all proposed CIM solutions in the seme context and compare them thoroughly in terms of power consumption, speed, are and reliability. This lack of comparative studies makes it difficult to decide which is the right technology and CIM implementation in order to maximize the performance of specific application. In this dynamic context, the choice of technology and computing paradigm is driven by the personal experience of the circuit designer rather than performance optimization.

The purpose of this project is to provide a tool (modular and expendable over different resistive memory devices and computing paradigms) which will guide the designer towards the optimal combination technology/computation for desired application. This goal will be achieved by developing and implementing a Computer Aided Design (CAD) platform to be integrated with standard CAD tools to support introducing resistive memory devices into the standard integrated circuit design flow. More precisely, the following objectives will be fulfilled: (O1) Create a data-base of primitive gates for computing-in memory for different types of resistive-memory devices and arrays; (O2) Develop an automated tool for design space exploration and circuit analysis; (O3) Develop an automated tool to map critical computation kernels (specific to application) on appropriate compute primitive to achieve desired performance metrics.
Mis à jour le 2 March 2022