Addressing System Integration Challenges for Processing-in-Memory Architectures
Christina Giannoula
Max Planck Institute for Software Systems
02 Dec 2026, 12:15 pm - 1:15 pm
Saarbrücken building E1 5, room 002
Joint Lecture Series
Modern applications, including machine learning (ML), bioinformatics, and
databases, process increasingly large volumes of data. This data growth creates
memory bandwidth bottlenecks when moving data between memory and processors.
This trend has motivated a decade of research on computer architectures that
move beyond the limitations of the von Neumann model and toward truly
data-centric designs. Processing-in-Memory (PIM), also referred to as
Processing-Near-Memory (PNM), is an emerging computing paradigm that mitigates
this data movement bottleneck by integrating processing units within modules
(e.g., ...
Modern applications, including machine learning (ML), bioinformatics, and
databases, process increasingly large volumes of data. This data growth creates
memory bandwidth bottlenecks when moving data between memory and processors.
This trend has motivated a decade of research on computer architectures that
move beyond the limitations of the von Neumann model and toward truly
data-centric designs. Processing-in-Memory (PIM), also referred to as
Processing-Near-Memory (PNM), is an emerging computing paradigm that mitigates
this data movement bottleneck by integrating processing units within modules
(e.g., DRAM), so that computation takes place where the application data
resides. However, fully exploiting the benefits of PIM hardware requires
solving key systems challenges in correctness, performance, and programmability.
In this talk, I will present my research on designing PIM systems that address these challenges. After an overview of application-driven PIM design, I will discuss two systems challenges on two recent PIM architectures. First, I will describe how we solved the synchronization problem on PIM architectures with a shared memory model, providing both correctness and high performance on state-of-the-art applications. Second, I will present our ongoing work on the programmability problem on PIM architectures with a distributed memory model, where we automate kernel deployment for emerging ML models. I will conclude with current trends in the hardware design of PIM-based architectures and the new systems challenges they raise. I will illustrate that cross-stack integration is essential for PIM architectures to achieve widespread adoption.
Read more
In this talk, I will present my research on designing PIM systems that address these challenges. After an overview of application-driven PIM design, I will discuss two systems challenges on two recent PIM architectures. First, I will describe how we solved the synchronization problem on PIM architectures with a shared memory model, providing both correctness and high performance on state-of-the-art applications. Second, I will present our ongoing work on the programmability problem on PIM architectures with a distributed memory model, where we automate kernel deployment for emerging ML models. I will conclude with current trends in the hardware design of PIM-based architectures and the new systems challenges they raise. I will illustrate that cross-stack integration is essential for PIM architectures to achieve widespread adoption.