Research Projects

Modern computing systems are based on von Neumann architectures, i.e., the computing systems consist of separate compute (CPU/GPU cores) and memory elements. Since the memory and the computing cores are separate, each off-chip memory (DRAM) access by the computing cores consumes significant latency and energy. Hence, any memory intensive application suffers from poor performance while executing on von Neumann-based computing systems. In fact, recently emerged machine learning (ML) applications are both compute and memory intensive which makes traditional von Neumann-based systems a bad candidate to execute ML applications. To address the challenge of energy inefficiency occurring from off-chip memory access, researchers have developed in-memory computing (IMC) technology. The elements in IMC-based systems act as storage as well as processing elements eliminating the need for an off-chip (separate) DRAM. IMC-based systems have been proven to deliver high energy-efficiency while performing matrix-vector multiplication (MVM). Since most of the ML applications mainly consist of MVM, IMC-based system is an excellent candidate to execute ML applications in an energy-efficient manner. Till date, several researchers proposed various architectures for IMC-based systems to execute ML applications. The aim of most of the existing IMC-based ML accelerators are to integrate the IMC devices to ensure high energy-efficiency for a particular class of ML application (e.g., deep nerual network (DNN), graph convolutional network (GCN)). However, the scope of the computing requirements now-a-days is not limited to a single class of ML application. Therefore, IMC-based ML accelerators tailored for a single ML application are not entirely suitable for real time deployment. Indeed, executing different types of ML applications on a single IMC-based accelerator is challenging since different class of ML applications exhibit different computing and data access pattern. Specifically, there are two main challenges when multiple class of ML applications are being executed on a single system - 1) Mapping: which part of the ML application will be executed on which part of the system, 2) Scheduling: when a part of the ML application will be executed. None of the existing IMC-based ML accelerators address these challenges independent of class of ML applications. Moreover, any computing system also needs an operating system (OS) which provides the user an interface to program the computer. There exist no dedicated OS for IMC-based ML accelerators. To this end, we propose INERTIA; IMC-based energy-efficient unified (independent of ML applications) ML accelerator. The proposal contains a fast and accurate simulator to enable quick design space exploration; ML application-agnostic energy-efficient mapping and scheduling technique; a software ecosystem to be provided to the end users to make use of the proposed IMC-based ML accelerators.