|
DYNAMIC LOGIC SYNTHESIS
Abstract
I will describe our research plan to develop automatic logic synthesis (and
layout) capabilities for random logic blocks in which the majority of the
logic is implemented with dynamic CMOS gates. We are currently developing
the four dynamic logic synthesis and layout tools listed below. Although
each of the synthesis tools will be aimed primarily at random (control)
logic synthesis, the methods could be used for datapath as well.
Clock-delayed (CD) domino
This is a new single-rail dynamic logic family.
We are currently developing a new technology mapping algorithm
that will utilize more complex gates than in our current approach.
No pull-down path greater than 2 transistors will be permitted.
Tremendous speed ups over static logic have been obtained (2.2 to
more than 6 times faster) since considerably fewer logic levels
are needed and each gate is really fast since it's limited to
pull-down chains of no more than 2. Sun Microsystems is currently
using our CD domino synthesis tool.
Mixed dynamic-static logic
In this approach, the random logic description is first minimized
and made positive unate (just as if it were dual-rail domino). We
then apply a new technology mapping algorithm (which turns out to
be optimal) that ensures that each dynamic gate (or cell) will have
no pull-down path greater than 2 transistors and that each static
gate (cell) will have no pull-up path greater than 2 transistors. Then
we must make the network "two colorable" (i.e. a dynamic gate can only
drive a static and vice versa). The algorithms for these steps are
quite interesting. This will be at least as fast as dual-rail domino,
but will use considerably less area and power. Conventional domino
logic clocking is employed.
Decomposition into two-level logic structures
We are developing a novel tool which takes a minimized random logic
description and automatically decomposes into multiple area-efficient
two-gate-level logic structures. Each two-level structure is extremely
fast and compact since it is implemented as a CD domino NOR-OR
structure. Not only is it two gate levels, but each "level" has a
pull down path of one logic transistor for maximum speed. We have
observed tremendous speed ups over static CMOS for the MCNC
benchmark set.
Self-resetting dynamic logic
In this synthesis tool, the random logic is first minimized (using
Synopsys or SIS) and then is passed on to our novel technology
mapping algorithm which results in a decomposition into gates
having a longest pull down path of two logic transistors. The
resulting netlist is then levelized. Utilizing a variant of the CD
domino technique, each level (or stage) is made self-resetting.
A stage goes into precharge as soon as it has finished evaluating AND
the succeeding stage has finished evaluating. Also, a stage enters
evaluation as soon as it has finished precharging AND the previous
stage has finished evaluating. This tool is particularly promising
for pipeling applications. The latency is two stage delays (keep in
mind that each stage delay is quite short, since each gate has a
maximum pull-down chain of 2 devices). Note that this approach to
pipelining requires no explicit latches and no signals must be
propagated through latches. The self-resetting tool can also be
used to synthesize very fast non-pipelined circuits, in fact,
much faster than static CMOS.
|