Pipelined Virtual-Channel-Based Routers - Microarchitecture of Network-on-Chip Routers - page 163

0 1 2 3 4 5 6 7 8

VA
su

cc

H0 LT-BW RC-EB SA-DQ LT-BW

ST

cc

B0 LT-BW EB SA-DQ LT-BW

ST

cc

T0 LT-BW EB SA-DQ LT-BW

ST

VA
su

cc

H1 LT-BW RC-EB SA-DQ LT-BW

ST

Fig. 9.19 An example of the operation of the 4-stage pipelined router RC|VA|SA|ST for the flits
of two packets that arrive back-to-back on VC#0 and VC#1 respectively

When the two input packets arrive on different input VCs, the first packet belongs
to VC#0 and the second one to VC#1, the head flit of the second packet is at the
frontmost position of the input VC#1 buffer in cycle 3, according to Fig. 9.19 ,and
not in cycle 4 as done in the previous case. Therefore, it is allowed to complete RC
and move to the intermediate EB of VC#1 one cycle earlier. This input traffic allows
in cycle 4 the concurrent use of all stages of the pipelined VC-based router. The
head flit of the first packet is in ST and moves to the selected output port; the body
flit performs SA, and the tail flit of VC#0 and the head flit of VC#1 are enqueued
on the corresponding EBs. Then, in cycle 5, while the tail flit of the first packet
participates in SA and wins, the head flit of the second packet is free to participate
to VA and allocate an output VC to its destined output port.

9.6 Take-Away Points

The operation of VC-based routers can be pipelined based on three primitive
pipelined organizations that separate the execution of the corresponding task RC,
VA, or SA from the ones that follow. Using such primitive pipeline stages, and
following a compositional approach that stitches together the primitive pipeline
stages for RC, VA and SA, multiple alternatives can be derived for the design
of deeper router pipelines. The derived designs achieve a high operating clock
frequency and overlap in time as much as possible the execution of the tasks required
per packet and per flit. The placement of pipeline registers can be done either in
the control path or in both the control and the datapath, depending on the chosen
organization and the number of idle cycles that can be sustained in the flow of flits
inside the router's pipeline.