( ESNUG 483 Item 6 ) -------------------------------------------- [11/19/09]
From: [ Snarf of the Thundercats ]
Subject: One user's benchmarks of Magma Talus from 1.0.86 to 1.0.91-beta
Hi John,
Please keep me anon on this post.
I've recently been working with the new Talus beta release from Magma and
thought I'd share with you some of my observations. The most noticeable
change is an improvement in closing designs for TNS. These numbers track
the same design that I've pushed through 4 different releases of the tool,
including a recent beta release.
Build Worst-Slack Total-Slack No. of violations
Talus 1.0.86 -178p -615n 9.0K
Talus 1.0.89 -175p -430n 7.5K
Talus 1.0.91-beta -100p -70n 3.0K
This improvement is very important for us. Our designs usually go through
many architectural changes. We do not have the luxury of time to get used
to the design, and may not have the best possible floorplan and our design
may not close timing completely. If Talus can optimize our design in this
scenario it leads to an overall TAT improvement
The other metrics I've seen are better slew fixing and reduced congestion.
Slew Fixes
Build Number of Violations
Talus 1.0.86 1890 violations
Talus 1.0.88 200 violations
Talus 1.0.91-beta 43 violations
Congestion measured at placement
Build Percent of buckets that overflow
Talus 1.0.89 3.5%
Talus 1.0.91-beta 2.0%
For the overall flow, as I understand it, Magma R&D worked to develop
several simplified flows (they are calling them "out of the box") that
reduce the number and complexity of commands to drive the tool. The
starting point is good, but as you can imagine, we have customized them
for our own needs. Even though we added our own tweaks, it is still way
better than before and easier to drive. These reference flows let
us spend time customizing to add value for our designs, not doing basic
flow-engineering.
ECO Routing in this new release has improved as well - we have seen an
average 3x improvement in ECO convergence runtime. Some comparative ECO
routing runtimes that we are trying to deploy in our flow with the help
of the AEs:
Build ECO Routing Runtimes
Talus 1.0.89 6 hours
Talus 1.0.91-beta 1.2 hours
Overall runtimes through the full flow have vastly improved in the beta
version - we have seen throughput go up by a factor of 1.5x for full
placement and optimization.
One thing that I like about Magma is the way they have wrapped their entire
flow around a single data model plus the flexibility we have in driving it
via TCL. Taking in late changes is never fun, but at least in the Magma
system they are relatively easy to take care of since there are not multiple
tools and databases, but just the one data model. Recently we were far
along in one of our designs when we discovered that we had been using the
wrong library cells. It was trivial to swap in the new cells and keep
moving forward. In the Cadence/Synopsys flows I've worked with, this would
have thrown a major delay into the schedule.
Gotchas:
- One problem area for Talus 1.0.91-beta is in clock tree synthesis. We
would like to see better insertion delay and skew consistently across
all of our designs.
- We would also like more control over CTS; specifically we'd like to be
able to influence routing width/spacing for clock nets, and buffer
selection, clock area, power, etc.
- Additional work is needed on DRC cleanliness. Talus has improved from
the last release, but still needs some additional work.
Regardless, we were pleasantly surprised at the results we've seen so far
with this beta Talus release.
- [ Snarf of the Thundercats ]
Join
Index
Next->Item
|
|
