The personalized recommendation system's continuous size growth poses new challenges for model inference. Although weight-sharing algorithms have been proposed to reduce embedding table capacity, they increase memory access. Recent advancements in processing-in-memory (PIM) successfully enhance the recommendation system's throughput by exploiting memory parallelism, but our analysis shows that those algorithms introduce CPU-PIM communication overhead into prior PIM systems, compromising the PIM throughput. We propose ProactivePIM, a specialized memory architecture integrated with PIM technology tailored to accelerate the weight-sharing algorithms. ProacitvePIM integrates an SRAM cache within the PIM with an efficient prefetching scheme to leverage a unique locality of the algorithm and eliminate CPU-PIM communication.

翻译：个性化推荐系统规模的持续增长给模型推理带来了新的挑战。尽管已有权重共享算法被提出以减少嵌入表容量，但它们增加了内存访问量。近期存内处理（PIM）技术的进展通过利用内存并行性成功提升了推荐系统的吞吐量，但我们的分析表明，这些算法在先前的PIM系统中引入了CPU-PIM通信开销，从而影响了PIM的吞吐性能。我们提出ProactivePIM，一种与PIM技术集成的专用内存架构，旨在加速权重共享算法。ProactivePIM在PIM内部集成SRAM缓存，并采用高效的预取方案，以利用该算法特有的局部性并消除CPU-PIM通信。