( DAC 00 Item 46 ) --------------------------------------------- [ 7/13/00 ]
Subject: Simplex, Mentor xCalibre, Cadence HyperExtract, Sequence, Avanti
PARACITIC EXTRACTION: One very crowded house here. Seems like everyone has
some sort of 3D or 2-1/2D extractor to sell. Looks like this year and next
some consolidating is going to happen. Started with Sente merging with
Frequency. Lots of others are reselling Random's QuickCap, too.
"Parasitic Extraction
The Holy Grail of accuracy has always been the Raphael field solver,
which was recently acquired by Avanti when they bought TMA. It is super
accurate but takes enormous amounts of computer time; you can only run
it on a couple of critical nets. Random Logic Corp. (RLC - get it) has
created another field solver called QuickCap. It is quick ONLY in
relation to Raphael - it's still mighty slow. It's faster than Raphael,
but gives more limited information. For example, it only provides the
total net capacitance; it doesn't break it down into segments. A slew
of other companies plan to resell RLC's technology as part of their
tool. Avanti has wrapped a GUI around it and is now reselling it,
apparently as a new "mode" of Raphael. Simplex and Sequence (formerly
Frequency Technology) will both add it to their extraction tools, and
Monterey Design Automation uses it as part of their tools suite.
The Simplex salesman seemed the most confident - he claimed that his
tool was much, much faster than Sequence (formerly Frequency Technology)
or Mentor's xCalibre, although he admitted that Sequence's extraction
tool was a bit more accurate. Interestingly, both Sequence and Mentor
claim their next release will include revolutionary changes to speed up
their tools. Simplex claimed the best validation in silicon of any
extraction tool. He did seem a bit afraid of Avanti, but noted that his
tool worked directly from the DEF, which was the reason for its high
speed, while Avanti has to translate the design into an internal format.
They say that about 90% of their nets will be within 10%, and the rest
are within 15%. No nets are more than 15% off.
Frequency Technology and Sente have merged to form Sequence. Their
salesman admitted that it is currently much slower than Simplex, but
said that they are rewriting the tool to directly use DEF (like Simplex
does) at which point it will be just as fast as Simplex but more
accurate. The new release is due in July. He said it was the only 3D
extraction with inductance. It can also split jobs between multiple
processors. He said total capacitance was within 5% of a field solver
and coupling capacitance was within 10%.
Mentor's salesman admitted that xCalibre was not as fast a Simplex, but
said that they will be releasing a hierarchical version in Q3 that will
be much faster. He said that 80-90% of the nets are within 10% of a
field solver. He said that the most inaccurate nets are short ones
with weird topologies, and noted that these don't matter as much, since
on short nets most of the net capacitance in gate capacitance anyway;
the interconnect doesn't influence things as much. Interestingly, the
Cadence (formerly Lucent) AE countered this by pointing out that both
analog designs and test structures often contain short runs that where
accurate capacitance is crucial.
Cadence's Assura RCX is the "Clover" tool they bought from Lucent. They
say it is slower but more accurate than Cadence's Hyperextract, and it
is not clear if both tools will survive indefinitely. They say it is
within 10% of a field solver, although they correlated it to Lucent's
TLP field solver, rather than Raphael. In general, the former Lucent
people seemed to be from an isolated environment and didn't quite know
how they ranked compared to the competition. The AE emphasized that the
accuracy was not dependent on the length of the net. It sounded like it
was the best tool for doing odd jobs like analog designs, arbitrary
shapes, air gaps, etc. It currently extracts 20K-30K nets per hour at
the transistor level, more at the cell level. It can currently do
either flat extraction or a type of hierarchical where the routing sees
the cells as a GND plane and the cells don't see the routing at all. A
true hierarchical extractor is in the works. It can extract C only,
lumped RC or distributed RC. L is coming next year. One interesting
feature is that it can extract a critical path from a design to allow
SPICE simulation, and it automatically ties the unused inputs to the
correct state to allow an edge to propagate down the path. It works
within dfII like Diva (it can read Diva rules, too) and is integrated
into Analog Artist, and currently uses the Pillar database. It is not
part of Cadence's new binary Genesis database so it's not clear how it
fits in the flow with PKS - it would probably be just a final check.
One interesting point the Cadence AE made that I assume applies to
everyone. The cladding on copper interconnect means that the
resistivity can vary depending on the width of the line (I don't
understand this, I'm just repeating it) so their tool doesn't handle
copper that well as of yet.
Avanti says their RCXT tool is 10X faster than their old RC tool and
allows for distributed processing. It works hierarchically and reads
LEF/DEF but translates it to an internal format (the Simplex salesman
called this a disadvantage). One very interesting feature is that it
can work on flawed layouts. It can understand that shorted power/gnd
nets are supposed to be separate and do an extraction as if they were,
and also that an open net was actually continuous. It has the Random
Logic Corp Quickcap field solver built in to run on selected nets.
Even though Avanti owns the industry standard Raphael field solver, it
is so slow that it can't be run at the chip level. RCXT uses Avanti's
common Milkyway database. It does clock skew analysis on each net but
has no automated way to solve problems. It can also do rail drop
analysis like Epic's Railmill tool. It doesn't have a simulator built
in - it uses either VCD from a simulation run done earlier, or an
estimated percent activity for each net (you can use wildcards when
specifying). It will soon accept the Synopsys SAIF (Switching Activity
Interchange Format) as well. Power per cell comes either from a
Synopsys .lib file or a SPICE netlist(?).
Silvaco has a hierarchical netlist extractor called Hypex that they say
is super fast, but they have no benchmark data. They also have their
own field solver that does distributed capacitance (unlike the Fastcap
tool). It currently does only RC; L is due in September. They also
have a tool for doing transmission line modeling (including skin
effects) one net at a time.
OEA International sells a field solver like Raphael; it does distributed
RLC. They also have a tool like Airmail to analyze ground bounce and
simultaneous switching.
Virgules sells a tool to characterize interconnect RC based on a process
description."
- an anon engineer
|
|
