A Full-System Perspective on UPMEM Performance

Recently, UPMEM has introduced the first commercially available processing in memory (PIM) platform. Its key feature are DRAM memory chips with built-in RISC CPUs for in-memory data processing. Naturally, this has sparked interest in the research community, which previously was limited to PIM simulators and custom FPGA prototypes. One result of this is the PrIM benchmark suite that combines an in-depth analysis of PIM performance with benchmarks that measure the speedup of PIM over processing on conventional CPUs and GPUs [10]. However, the current generation of UPMEM PIM faces limitations such as memory interleaving, and as such does not provide true in-memory computing. Applications must store data in DRAM and transfer it to/from UPMEM modules for processing, which behave just like computational offloading engines from this perspective. This paper examines the ramifications of treating them as such in comparative performance benchmarks. By extending the PrIM suite to address the challenges that computational offloading benchmarks face, we show that such a full-system perspective can drastically alter offloading recommendations, with 9 of 11 previously UPMEM-friendly benchmarks now performing best on a conventional server CPU.