In-memory data processing is the state-of-the-art approach for large-scale data analytics. For a more efficient processing and to be able to store more data in main memory, new hardware developments for memory such as Compute Express Link (CXL) and High-Bandwidth Memory (HBM) are of great interest. However, these developments also heterogenize the memory hierarchy, increasing the importance of efficient data transfers between different memory types and locations.
To this end, Intel offers a new on-chip memory-operation accelerator called Data Streaming Accelerator (DSA) in modern Xeon Scalable processors. In this paper, we introduce this accelerator and present results of an experimental analysis from the perspective of in-memory processing. Moreover, we draw conclusions from our experimental analysis and finally discuss a concrete optimization idea for the in-memory data processing using DSA. Our idea is called on-the-fly data distribution (ODD) and considers HBM as a memory accelerator for DRAM-oriented in-memory data porcessing. Our evaluation shows that the DSA increases the benefit of ODD compared to a pure CPU-based approach.