(PhD02) Scaling OpenPOWER Near-Memory Computing Architecture
TimeMonday, June 17th1pm - 6pm
DescriptionHPC system designers are moving away from the conventional compute-centric model towards a more data-centric approach, in which data movement is reduced by implementing processing capabilities close to where the data resides in the memory system. With "close" being a relative term, there is a wide range of possibilities to bring computation nearer to the data, resulting in various architectures being investigated today. This type of approach is denoted as near-memory computing or processing.
Near-memory computing is one of the few real solutions to address the current scaling issues in HPC systems to realize exascale computers that are needed for near-future Big data workloads. However, near-memory computing is still in its infancy. Before it can be established as an essential component of HPC systems and be exploited for accelerating Big data workloads, multiple challenges have to be addressed. Besides the design of the near-memory computing device itself, this also includes its integration into the overall computer system architecture, how multiple near-memory computing devices can work together to scale to larger data volumes.
We developed an NMC framework using ensemble learning that combines hardware parameters and application-specific characteristics and provides performance estimation of new applications and different system designs. This framework can provide rapid exploration compared to a state-of-the-art NMC simulator. Our current focus is on intranode scaling for weather forecasting application on POWER9 and HBM with near-memory accelerators via high bandwidth OpenCAPI interface.