The exponential growth in digital data has fueled the emergence of data-intensive applications like large language models, digital twins, and metaverse. These applications have unprecedented computing demands that exceed the capabilities of today's high-performance computing systems. To address this challenge, new computing paradigms are being explored. One promising solution is to perform processing in memory (PIM) using emerging hardware. However, PIM paradigms are still in their infancy. A major obstacle to the widespread adoption of PIM paradigms is the lack of advanced electronic design automation (EDA)-based designs to ensure the system is both robust and energy-efficient.
This dissertation presents electronic design automation-based solutions to enhance the robustness, scalability, and efficiency of PIM systems. The initial phase of this research explores the combination of different PIM paradigms, aiming to leverage their strengths to strike a balance between high-assurance computation and resource constraints. This involves exploring the combination of analog-digital and digital-digital computing paradigms. In the subsequent phase, advanced computer-aided design (CAD) techniques are employed to optimize the mapping of computation to in-memory platforms. This includes developing automated synthesis flows to decompose computations into in-memory computation kernels and creating hardware-efficient execution sequences to maximize the computational efficiency of PIM systems. In the final phase of this dissertation, novel architecture-level solutions for PIM systems are developed to enable massive parallelism for data-intensive computations while minimizing cross-architecture data transfer costs.
The research projects in this dissertation aim to facilitate the advancement of in-memory computing paradigms to accelerate modern data-intensive applications.
Rickard Ewetz, Chair.
Read More