Abstract:

Many modern and critical workloads, such as machine learning, computational biology, graph processing, databases, video analytics, and real-time data analytics, face significant challenges due to the bottleneck caused by data movement. These workloads typically involve irregular memory accesses, limited data reuse, inefficient cache line utilization, low arithmetic intensity (i.e., the ratio of operations per accessed byte), and large datasets that far exceed the size of the main memory. The data movement bottleneck severely constrains the energy efficiency and performance of modern computing systems. This bottleneck results from the conventional Von-Neumann architecture, which separates the processing unit from the data storage location. To address this issue, we need to consider replacing the conventional processor-centric design with a more data-centric approach. That is, computations are done closer to or integrated within the data storage units, instead of within the processing unit. This approach, known as Processing-in-Memory (PIM), aims to mitigate the challenges posed by data movement. In this talk, we aim to raise awareness about the concept of PIM and present a few examples of solutions we designed with real PIM products.

Bio: