Mont-Blanc 2020 partner Benoit Dupont de Dinechin (CTO, Kalray) is an invited speaker of the Supercomputing Frontiers Europe 2018 conference. He will share his vision on “Manycore Accelerators beyond GPU Architecture”.
High-end GPU processors implement a throughput-oriented architecture that has been highly successful for CPU acceleration in supercomputers and datacentres. GPUs can be characterized as manycore processors, that is, parallel processors where a large number of cores are distributed across compute units. Each GPU compute unit or «streaming multiprocessor» is composed of multi-threaded processing cores sharing a control unit, a local memory and a global memory hierarchy. However effective, GPU architecture entails significant limitations in the areas of expressiveness of programming environments, effectiveness on diverging parallel computations, and execution time predictability.
We discuss the architectural options and programming models of accelerators designed to address high-performance applications ranging from embedded to extreme computing. Similarly to GPU architectures, such accelerators comprise multiple compute units connected by on-chip global fabrics to external memory systems and network interfaces. Selecting compute units composed of fully programmable cores, coprocessors and asynchronous data transfer engines enable to match the acceleration performance and energy efficiency of GPU processors, while avoiding their limitations. This discussion is illustrated by the co-design of the 3rd-generation MPPA manycore processor for automated driving, and is related to the implication of Kalray in the Mont-Blanc 2020 and the European Processor Initiative projects that target exascale computing.