Editorial for the special issue on heterogenous computing

Abstract

shows that the Ascend 910 AI chip can achieve up to 216 TFLOPs for float16 data for the matrix unit and 3390 GFLOPS for the vector unit. • The paper written by Shen et al. (2024) proposes a resistive random accessed memory (ReRAM) based processing-in-memory (PIM) accelerator named ReGCNR for GCN-based recommendation. It fits with large-size embedding table and user-item graph with 3-dimensional (3-D) stacked heterogeneous ReRAM and maximizes the efficiency of the execution pipeline using a joint degree mapping schema. The performance can be improved by assembling a well-coordinated pipeline and hardware scheduling design.The second part of special issue focuses on the big data applications with heterogenous computing, including graph data processing, streaming data processing, data compression, non-uniform data sampling.