Next: SSA in buffers
Up: Reducing Spurious Switching Activity
Previous: Overhead for gating
Three output network topologies were introduced in
Section III. If all the output data are sent to all
the receivers without regard to whether they need the data or not,
the MST output network will consume the least power and the fully
dedicated the most. If interconnect SSA is suppressed using the
proposed techniques, however, this will not always be true.
Fig. 10 shows a DPU, colored black, sending
data to two other DPUs, colored white, using different output
network topologies. It also shows the optimal locations for gating
signals with SSA as small triangles.
Figure 10:
Different output network topologies for
(a) one DPU sending data to two other
DPUs:
(b) MST,
(c) fully dedicated, and
(d) trunk-branches.
|
Suppose tri-state buffers are used so that the last useful value
is remembered in the gated idle cycles. Let denote the
unit-length switched capacitance for data, say, for variables
, needed by the upper receiver, which is estimated as
where
is the unit-length switched capacitance in
the wire when is followed by . That for the lower
receiver, , and that for all the output variables of the
sender, , can be estimated in the same way. Assuming
signal gating is performed at the optimal locations shown in
Fig. 10, we can estimate the total wire
switched capacitance for the output network implemented as
follows: MST, , fully dedicated, , and
trunk-branches, , as
The above equations imply that
Therefore, when
, we have
and when
,
.
Since
can be either larger or
smaller than , none of these topologies
is always better than the others in terms of power
consumption. In this work, the trunk-branches topology
is assumed due to its never-the-worst power consumption
and the algorithmic simplicity for constructing it.
Next: SSA in buffers
Up: Reducing Spurious Switching Activity
Previous: Overhead for gating
Lin Zhong
2003-10-11