Embedding Compute within NoC Channels

点击翻译按钮获取中文摘要

Contemporary computers connect many CPUs to a shared RAM, allowing them to co-operatively process the data. Shared RAM bandwidth is now a huge bottleneck, despite the addition of caches and wider data-paths. Distributed systems connect multi-core CPUs over networks but introduce new packet-batching overheads.POETS is a recently finished EPSRC programme grant which re-thinks computing in terms of events, using tightly coupled messaging and computation. Many applications have been developed which used the POETS methodology, along with a bespoke hardware fabric called Tinsel. POETS research is ongoing, and some remaining challenges are: 1) how can off-the-shelf compute hardware be used to make the results more accessible, and 2) how can processing be more tightly integrated into the network?This project builds on the large POETS research base and explores how to embed processing directly into a Network-on-Chip. The core idea is that when a processor sends a message to another processor the message can be transformed in transit. Such a system turns the network into an active computational device, rather than just a data movement layer. Challenges to solve include: how do we choose which functions to embed in the network? How do we prove the augmented network still performs the same function? How do we balance resource utilisation between smart network IP vs more processors? This research leverages ENNUI, a cutting-edge £250K hardware platform for exploring POETS-inspired architectures. The research aligns with HPC users and providers (such as ARM), with the ultimate impact being that it reduces power consumption and improves sustainability for large-scale compute.