( ESNUG 477 Item 6 ) -------------------------------------------- [11/20/08]
From: [ Dr. Evil ]
Subject: Why we dumped PrimeTime (and skipped Cadence ETS) for GoldTime
Hi, John,
EDA vendor politics require that I be anon.
We taped-out a 40 M gate design using Extreme DA's GoldTime Static Timing
Analyzer. Our chip is an array of processing elements with over 11 million
instances, fabricated using the 90 nm process technology.
With a flat design netlist and hierarchical parasitic file, GoldTime took
only 1.5 hours to compute and generate all the timing reports. PrimeTime
was taking around 3 hours for similar analysis, whereas the Cadence ETS
was over 6 hours.
By analyzing the design flat, we were able to optimize global signals
(clocks, resets and test modes) more accurately. Since GoldTime supports
hierarchical parasitic annotation, we further reduced our run-time by doing
smaller extractions of the design in parallel. PrimeTime can read in our
design flat but its run time was much longer. Time taken to read our
netlist and parasitic (SPEF) was comparable, but PrimeTime was much slower
in processing timing exceptions.
We began evaluating GoldTime static timing analysis tool in June 2006, and
purchased it a few months later based primarily on run-time, accuracy and
advanced clock skew analysis. The tool enabled us to reduce global clock
power and fix timing violations. We ran GoldTime against Cadence ETS
(Encounter Timing System) timing tool with less than 2% deviation. GoldTime
results were also within 2-4% of Synopsys PrimeTime results.
What GoldTime Did:
- Signal Integrity. We did all our timing analysis with GoldTime signal
integrity turned on. (i.e. GT 1.5 hr, PT 3.0 hr, ETS 6.0 hrs was with
SI.) I was impressed by how little time GoldTime took to perform the
full-chip, SI-aware timing analysis with propagated clocks.
- GoldTime is TCL-based and supports the SDC format, it easily integrates
into our design flow.
- Compatibility with Synopsys PrimeTime scripts. We use IP for DDR
controllers (etc.) which come with standard PrimeTime scripts. GoldTime
can process them 'as is', saving us weeks and possibly months. Not only
are the scripts compatible, GoldTime also uses the same SDC format for
timing constraints. It could take weeks to convert complicated Synopsys
PrimeTime TCL scripts to another timing engine. The timing report format
is also similar to PrimeTime.
- We used its clock skew analysis to improve our peak power by making our
clock skew tighter. Pre-optimization clock skew was over 200 psec.
After analysis and fixes to the clock tree, we were able to reduce it
around 80 psec with a target frequency of 500 MHz (2 nsec clock period).
Liked best:
1. GoldTime's ability to handle large designs with short run-times.
2. GoldTime's small memory footprint. For the design, we were able to load
entire netlist and parasites with less than 10 GB of physical memory.
(The Extreme-DA people worked tirelessly with us to reduce memory
footprint. We were barely fitting the full chip flat -- so Extreme DA
bought their own 16 GB machine and made sure the test cases all fit
on that size machine.) We didn't do a PT memory footprint analysis,
but Cadence ETS analysis exceeded 16 GB.
Needs work:
1. Extreme DA could offer more enhanced SDC compaction techniques to reduce
our 1.5 hour runtime even further.
2. Add a .lib generation feature. Although GoldTime has all the info it
needs to write out a .lib, it doesn't support auto-.lib generation.
Using another tool to generate .lib models and incorporating it back
into the STA flow is not the best practice.
Extreme's R&D had unbelievably fast turnaround - we got same day responses to
many requests. They were able to accommodate all our unique requirements
like hierarchical query to the timing db, the ability to manage clock skews for
power and timing, and the smaller memory footprint.
I highly recommend GoldTime, especially to people using a PrimeTime flow.
It's simply better than PrimeTime.
- [ Dr. Evil ]
Index
Next->Item
|
|
