High-Speed Allocators for VC-Based Routers - Microarchitecture of Network-on-Chip Routers - page 140

8.5 VC-Based Routers with Input Speedup

Allocation efficiency can be improved by letting more than one VCs per input to
reach independently the crossbar thus allowing the switch to offer speedup at the
input side. 1 In the baseline organization of a VC-based router discussed so far, there
is only one input to the crossbar per input port, and thus only one virtual channel in
an input port can transmit a flit in a cycle. Allowing more than one VCs per input to
reach the crossbar directly, means that the inputs of the crossbar are multiplied with
the number of transmit ports per input. The input VCs can be separated in m groups
where each group receives a dedicated input port of the crossbar. This organization
is depicted in Fig. 8.6 . In this case the one per-input multiplexer that switched V
input VCs is replaced by m smaller multiplexers that select between V=m input
VCs. Equivalently, the input port that used only one input of crossbar, now sees m
inputs available. The output multiplexers of the router grow from N inputs to m N
inputs, since now every input is allowed to send flits from m different VCs assuming
that they move to different outputs.
Virtual channel allocation does not need to change, unless it is simplified to treat
the group input VCs as virtual networks and thus limiting the set of output VCs
that each input VC can allocate. On the contrary, switch allocation should change in
order to support the input speedup of the VC-based routers. In this case, SA1 that
operated using a V W 1 arbiter per input, now should support m arbiters that run in
parallel each one receiving V=m requests. At the output side the number of SA2
arbiters do not change and remains equal to one per output. However, since now
each input may receive the flits from the m N different inputs of the crossbar,
each output arbiter should handle m N requests. In overall, the delay of the SA
is not expected to change since the arbiter's delay depends logarithmically on the

Input #0

Output #0

Fig. 8.6 Input speedup
increases the input ports of
the crossbar thus allowing
multiple input VCs of the
same input to allocate an
output port and move to their
selected output. This
organization partially
alleviates the problems
inherent in separable switch
allocation thus increasing the
overall throughput of the
router

m

mNxN
crossbar

Input #N-1

Output #N-1

1 Input speedup means that although the rate of incoming flits is one flit per cycle for each input the
rate of the flits that can leave each input towards the crossbar is larger than one.