PipelinedWormhole Routers - Microarchitecture of Network-on-Chip Routers - page 84

credit
updat e

outPort outAvailable[0...N-1]

ready out[0...N-1]

CC

RC

en

SU

req

1

outLock 1

SA

dst

en

head

credit update

pipeline
registers

valid
data

valid

ST

data

Input #i Output #j

Fig. 5.13 The organization of the router that pipelines RC-SA from ST using a pipeline register
both in the control path, that registers the grants of SA, and in the data path, that registers the data
arriving at the input of the crossbar

0 1 2 3 4 5 6

cc
RC-SA-DQ

H LT-BW ST LT-BW

su

cc
s SA-DQ

B LT-BW ST LT-BW

cc
s SA-DQ

T LT-BW ST LT-BW

cc
RC-SA-DQ

H LT-BW ST LT-BW

su

Fig. 5.14 The operation of a pipelined router that executes RC-SA in the first pipeline stage and
ST in the second, and uses pipeline registers both in the control path and in the datapath of the
router

allowed between the operations per pipeline stage is fully revealed. The body flit is
on the link, the tail flit performs ST, while the head flit of the next incoming packet
from the same input (it could be also from a different input), performs RC and SA
and at the end of the cycle moves to the data pipeline register after having consumed
an available credit.
The main contribution of the SA pipeline stage is the isolation of the crossbar
from the rest of the control and data path logic (the remaining data path logic
consists mostly of the data multiplexing inside the input buffers). Depending on
the radix of the router, its data width, and other placement options the crossbar may
have a significant contribution to the final delay. In low-radix cases, though, the
delay of the crossbar is not the critical factor that determines the speed of the router.
The idle cycles that appear in the straightforward SA pipelined organization
can be removed by either relying to a re-organization of the input request unit
that forwards the body and tail flits directly to ST, provided that they have credits
available, or by adding another data pipeline register placed at the inputs of the
crossbar that would align the arrival of grants and the corresponding data.