PROJECT TITLE :

Mitigating Memory-Induced Dark Silicon in Many-Accelerator Architectures

ABSTRACT:

Many-Accelerator (MA) systems are introduced as a promising architectural paradigm which will boost performance and improve power of general-purpose computing platforms. During this paper, we specialise in the problem of resource below-utilization, i.e. Dark Silicon, in FPGA-primarily based MA platforms. We show that except the sometimes expected peak power budget, on-chip memory resources kind a severe below-utilization issue in MA platforms, leading up to seventy five percent of dark silicon. Recognizing that static memory allocation-the de-facto mechanism supported by fashionable style techniques and synthesis tools-forms the most source of memory-induced Dark Silicon, we have a tendency to introduce a completely unique framework that extends conventional high level synthesis (HLS) with dynamic memory management (DMM) options, enabling accelerators to dynamically adapt their allocated memory to the runtime memory necessities, so maximizing the general accelerator count through effective sharing of FPGA's recollections resources. We show that our technique delivers important gains in FPGA's accelerators density, i.e. 3.eight×, and application throughput up to three.one× and 21.4× for shared and personal memory accelerators.