Realizing Energy-efficient NoCs With In-memory Routing

As the number of transistors integrated on the chip becomes larger and the number of functional modules continues to grow,the performance requirements for the No C also increase.As a key component of the No C,the router directly determines the power consumption and latency of the No C.Although traditional virtual routing has the advantage of low latency,it has more routing micro-components and consumes a lot of power when the number of routers in the No C grows along with the number of cores,so how to strictly control the energy consumption of routing has become an urgent problem for chip designers.In order to design a more energy-efficient No C,this paper proposes an In-memory routing design inspired by the PIM technology,which is divided into an In-memory routing scheme and an In-memory borrowed routing scheme.As a new routing design,In-memory routing relies on memory,and its internal structure contains only memory buffers,route computation units,and output port arbiters,while data transfer is realized through inter-memory dumping.In this paper,we propose an in-memory borrowing routing scheme based on the in-memory routing scheme.The difference between the two schemes is that the in-memory borrowing routing supports cross-port memory cell borrowing to cope with the congestion problem of heavily loaded ports(therefore imbalanced buffer utilizations).Both routing schemes are credit-based for traffic control.The credit value of the in-memory route is the number of free buffers in the memory cell,and the credit value of the in-memory borrowed route is the total number of free buffers within the route,and the credit-based traffic control mechanism is used to prevent memory buffer overflow.The experimental data shows that the No C based on in-memory borrowing route has improved limit injection rate compared to the No C based on virtual channel routing.Compared with traditional No C,the average latency of No C based on 6-port in-memory routing is partially reduced,and the average latency of No C based on 3-port in-memory routing is similar,while the average energy consumption of No C based on 3-port in-memory routing is reduced by about 78.6%～82.2%,and the average energy consumption of No C based on 6-port in-memory routing is reduced by about 81.3%～84.2%.