Merge Network for a Non-Von Neumann Accumulate Accelerator in a 3D Chip

Logic-memory integration helps mitigate the von Neumann bottleneck, and this has enabled a new class of architectures that helps accelerate graph analytics and operations on sparse data streams. These utilize merge networks as a key unit of computation. Such networks are highly parallel and their performance increases with tighter coupling between logic and memory when a bitonic algorithm is used. This paper presents energy-efficient on-chip network architectures for merging key-value pairs using both word-parallel and bit-serial paradigms. The proposed architectures are capable of merging two rows of high bandwidth memory (HBM)worth of data in a manner that is completely overlapped with the reading from and writing back to such a row. Furthermore, their energy consumption is about an order of magnitude lower when compared to a naive crossbar based design.