One key insight behind my activities in the MANGO H2020 project is accelerator architecture exploration. Among other developments, one central activity that I planned and I’m now leading is the development of a customizable manycore system based on a GPU-like organization, acting both as a tool for architectural exploration in ASIC development and as a deeply customizable FPGA overlay to be used on high-performance reconfigurable hardware platforms exploiting application-driven customization while still being programmable in pure software (a key prerequisite to ensure the actual usability of heterogeneous platforms in real environments). In fact, the GPU-like core, for now called nu+, will also come with a full compilation toolchain, based on LLVM, which will ultimately feature application-driven customization knobs. My long-term aim is to demonstrate the impact of the GPU-like concept in fields like scientific and parallel computing by porting demanding data-parallel applications and co-optimizing the nu+ hardware/software infrastructure in a deeply application-driven way.
(edit: A research proposal based on nu+ was recently granted the Intel Hardware Acceleration Research Program, HARP v2, offering early access to the new Broadwell + Arria10 FPGA MCP solutions provided by Intel to selected research centers worldwide)