Computational genomics plays an important role in health care, but is computationally challenging as most genomic applications use large data sets and are both computation-intensive and memory-intensive. Recent approaches with on-chip hardware accelerators can boost computing capability and energy efficiency, but are limited by the memory requirements of accelerators when processing workloads like computational genomics. In this paper we propose the accelerator-interposed memory (AIM) as a means of scalable and noninvasive near-memory acceleration. To avoid the high memory access latency and bandwidth limitation of CPU-side acceleration, we design accelerators as a separate package, called AIM module, and physically place an AIM module between each DRAM DIMM module and conventional memory bus network. Experimental results for genomic applications confirm the benefits of AIM. Due to the much lower memory access latency and scalable memory bandwidth, our non-invasive AIM achieves much better performance scalability than the CPU-side acceleration when the memory system scales up.
Congratulations again to Professor Cong, Professor Reinman, Fang, Gill, & Javadi!