The demands of the digital age have surpassed all guesstimates. This is apparent in the scores of datacenters that are trying their utmost to scale up to the burgeoning mass of data being fed into their servers. Cost overheads are escalating in terms of both server acquisition and energy consumption. We are almost at the brink, where existing infrastructures are unable to satisfy the needs of memory-hungry IT operations.
Many ongoing researches are fighting hard for a sustainable solution where demand can match supply. At the forefront of those researches is the development of specialized memory-centric architectures that can boost efficiency of memory utilization and minimize energy costs.
The reliance on in-memory processing to drive datacenter efficiency calls for two fundamental requisites of computing architectures: high memory resource utilization and cost-efficient access to memory. However, most datacenters still use processor-centric architectures that have long outlived the days of early IT computing. To break down this wall, the research proposes a “paradigm shift” toward specialized memory-centric architectures, where processing, networking, and software are all weaved around memory.
The research advocates the use of the developing die-stacked DRAM technology, which can boost memory access, reduce data movement, and thereby minimize energy overloads. The study has three important goals: the design of an innovative memory-centric architecture that can optimize bandwidth, power, and capacity, and result in a smaller energy footprint; the development of a system where processing can take place near the memory rather than in the CPU, which can potentially cause a dramatic drop in unnecessary data movement and consequently in the energy consumption; and scalable and low-cost virtual memory support to render high flexibility to near-memory accelerators.
The research is expected to provide novel solutions for datacenters to be in sync with the information flooding onto their servers. It could be a timely intervention, considering the fact that the volume of digital data is expected to multiply 300 times by 2030.