( DAC 11 Item 13 ) ---------------------------------------------- [10/13/11]
Subject: Extreme DA GoldTime vs. Magma Tekton vs. PrimeTime ECO
SLOW PRIMETIME: On the technical side, Extreme DA's GoldTime and their new
GoldX extractor has been publically embarrassing PrimeTime and Star-RC with
much faster STA and extraction results. Why? "Making 15 year old code go
faster is never easy," said Isadore Katz of CLK-DA. Magma Tekton is
also vying against the PrimeTime monopoly, but Extreme appears to be getting
the most user notice.
To counter the Extreme/Tekton threat, the PrimeTime folks revived the old
1997 failed "ECO Compiler" function in DC that was discontinued in 2001
and stuffed it back in as a part of PT now calling it "PrimeTime ECO". The
Synopsys DAC "PrimeTime SIG" was basically one long infomercial about how
super fantabulously great their "new" PrimeTime ECO is.
"What were the 3 or 4 most INTERESTING specific tools that
you saw at DAC this year? WHY where they interesting to you?"
---- ---- ---- ---- ---- ---- ----
We used GoldTime to sign off full-chips designs that ranged from
30 to 40 million instances at 40 nm and 28 nm. We've had 2 tapeouts
with GoldTime, and our silicon came back fully functional at speed.
Our GoldTime was able to handle 40 million instances in a full chip
run with no problem; its results correlated closely to SPICE.
In general, GT's Runtime is 2x faster compared to PT.
Tool Instances CPUs Runtime
GoldTime (GT) 40 million 16 24 hours
PrimeTime (PT) 40 million 16 Could not run full chip
GoldTime 1 million 4 1 hr 19 mins
PrimeTime 1 million 2 3 hr 19 mins
We used GoldTime for flat top-level STA sign-off. This approach gave us
far more confidence in the results over using PrimeTime and having to
divide up the chip into separate runs. GoldTime timing information is
normally within 3% compare to SPICE.
- The PrimeTime and GoldTime timing flows are similar, so we are able
to easily convert our PrimeTime scripts to GoldTime's script.
- GoldTime's signal integrity check is not exactly the same as
PrimeTime; GoldTime-SI's pushout (crosstalk delta delay) calculation
is a bit different from PrimeTime-SI's. GoldTime's method seems to
be more realistic in terms of the timing window between the
aggressors and victims.
- GoldTime's Clock Reconvergence Pessimism Removal (CRPR). Our chip
variation increased the pessimism in our timing estimates. CRPR
finds a common point in the clock tree, looks at the delay
differences, and removes this pessimism. It works well most of the
time. However, we did run into a bug where CRPR occasionally
mistraced and gave incorrect common clock points for our particular
high performance clock structure.
We also recently purchased Extreme DA's GoldX layout parasitic extraction
tool. GoldX's performance is at least 3x faster in runtime compared to
Synopsys Star-RC.
Tool Design size CPUs Runtime
GoldX 1 M instances 8 CPUs 16 mins
Star-RC 1 M instances 1 CPU 5 hrs
GoldX uses a built-in 3D field solver, which uses a random walk method.
98% of 3-sigma outliers were within 2% of our GoldX 3D field-solver
results.
Additionally, for timing signoff, GoldX SPEF and Synopsys STAR-RC SPEF
were within 2% of arrival time. Below is the extracted capacitance
difference between Star-RC and GoldX. Block1 to Block8 range in size
from 110k to 250k instances.
CTotal >1 fF >5 fF Arrival Time
Avg Diff % Avg Diff % Avg Diff % Compare %
block1 4.277 2.303 1.854 0.7103
block2 5.718 2.353 1.843 0.2188
block3 6.289 2.741 2.071 0.08142
block4 4.547 1.296 0.674 -0.2176
block5 5.748 2.196 1.888 0.3248
block6 6.210 2.476 2.328 0.3085
block7 10.751 2.373 1.939 -0.04674
block8 6.301 2.608 1.958 0.1521
Extreme-DA could improve their documentation/user guide, but their
support is very responsive.
Combining GoldTime and GoldX for our sign-off has dramatically reduced
our time to closure. We are able to turn around a full chip timing run
in only 24 hours for many last minute ECOs.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
a) Extreme DA/Atoptech showing POCV integration for a closed looped
PNR+Signoff flow. Interesting to see two small companies work
together to push a new technology.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
Timing Signoff @ 45nm/32nm - Looks like a lot of activity both on the
STA, SI and extraction front. In the past this area has been dominated
by Primetime-SI / STAR. Both tools have active competition now.
On the STA front, CLK-DA appears to have scaled way back, and Extreme-DA
has a bigger presence with a pretty busy booth.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
Extreme GoldTime looks like there are getting traction in the industry
touting POCV/SSTA and capacity, where PT is working on their capacity...
will be interesting to see if someone can break the PT monopoly.
On the extraction side Extreme announced GoldX. All the extractor
companies except Synopsys appear to have been rewriting their signoff
extractors. They all look in their infancy, but we'll see if someone
can finally get it right by designing an extractor targeting a multi-CPU
environment with the quirks of smaller geometries and variation.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
During DAC, I visited Magma's Tekton and QCP demos. The possibilities
of the tools were presented in good way. I got clear answers on all my
questions from the presenters.
Their comparison data to legacy tools (especially runtime) was based on
old versions of legacy tools, so it was not relevant.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
A) Synopsys Primetime ECO:
It's new for me to learn the Primetime ECO capability. It was first
brought to me by one of the new hired and then I saw the adoption of
Primetime ECO in other companies at DAC. I think it maybe helpful to
start evaluating this feature. Everything on paper looks good until
you actually use it. We will get Synopsys support and to see if it
can be useful in our design flow.
- [ An Anon Engineer ]
Join
Index
Next->Item
|
|
