Lawrence Berkeley National Laboratory

Many high performance applications run well below the peak arithmetic performance of the underlying machine, with inefficiencies often attributed to a lack of memory bandwidth. In this work we examine two emerging media processors designed to address the well-known gap between processor and memory performance, in the context of scientific computing. The VIRAM architecture uses novel PIM technology to combine embedded DRAM with a vector co-processor for exploiting its large bandwidth potential. The Imagine architecture, on the other hand, provides a stream-aware memory hierarchy to support the tremendous processing potential of the SIMD controlled VLIW clusters. First we develop a scalable synthetic probe that allows us to parametize key performance attributes of VIRAM and Imagine while capturing the performance crossover point of these architectures. Next we present results for two important scientific kernels each with a unique set of computational characteristics and memory access patterns. Our experiments isolate the set of application characteristics best suited for each architecture and show a promising direction towards interfacing leading-edge media processor technology with high-end scientific computations.