DAPPER: Data Aware Approximate NoC for GPGPU Architectures

High interconnect bandwidth is crucial to achieve better performance in many-core GPGPU architectures that execute highly data parallel applications. The parallel warps of threads running on shader cores generate a high volume of read requests to the main memory due to the limited availability of data cache space at the shader cores. This leads to scenarios with rapid arrival of reply data from the DRAM, which creates a bottleneck at memory controllers (MCs) that send reply packets back to the requesting cores over the NoC. Coping with such high volumes of data requires NoC architectures that possess high power overhead. To accomplish high bandwidth and low energy communication in GPGPUs, we propose Dapper, a data-aware approximate NoC architecture that increases the utilization of the available bandwidth by using low power single cycle overlay circuits for the reply traffic between MCs and shader cores. Dapper also incorporates a novel MC architecture that leverages the inherent approximability of the data values of certain applications and reduces the number of reply packets injected into the NoC by the MCs. Experimental results show that Dapper reduces the energy consumed in the GPGPU by up to 50% with up to 99% application output accuracy and minimum performance overheads compared to a state-of-the-art approximate NoC architectures.