( SNUG 04 Item 10 ) ---------------------------------------------- [08/11/04]
Subject: PrimeTime
PRIMETIME IS KING: While Synopsys is known in the investment community for
its dominance in the RTL synthesis market, most investors don't know about
the Synopsys virtual monopoly of the static timing market:
Dataquest FY 2002 Static Timing Analysis Market (in $ Millions)
Synopsys PrimeTime ######################################## $39.8 (97%)
Cadence Pearl # $1.1 (3%)
Mentor $0.1 (1%)
And on the technical side, Aart's bragging about his new rev of PrimeTime
being "3X faster" holds up -- or at that's what the users are saying.
5.) Aart said "PrimeTime 3X Faster!" in his speech. In your experience
is this real or B.S.? Do you use PrimeTime? Do you use 3rd party
PrimeTime add-ons like Prolific ProTiming and Sequence ShowTime?
Which one(s)? What do you think of them?
On PrimeTime I made a small test case.
Test conditions:
- large SoC that has taped-out
- 64-bit Solaris on identical servers with no memory limitation.
- PrimeTime versions compared are 2003.03-XXXXX and 2003.12-X.
Reading design and linking : 15 min vs 15 min => no improvement
Reading the constraints : 2 min vs 2 min => 8% improvement
2 timing updates : 120 min vs 40 min => 3.1X improvement
reporting : no figures => ~3X improvement
memory usage : 9 Gig vs 3 Gig => 3.2X improvement
I would then confirm, for our design, the 3X improvement that Aart is
talking about. For me the best of all is the memory usage drop (now
under 4 Gig) that would allow to run 32-bit PrimeTime under Linux to
gain some more (~30% memory and ~2X in runtime).
- Nicolas Verkinderen of Texas Instruments
This 3X for PrimeTime is real. Very real, and very much appreciated.
Our run times went down from 37 hours, 52 minutes using 2003.03 to
12 hours, 18 minutes using 2003.12. Memory requirements were improved
3X too, from 34 GB down to 11.2 GB.
- Steve McLafferty of Sandburst Corp.
PrimeTime is faster, no doubt about it. It absolutely screams on an
Opteron. But 3X? Compared to 2003.03, I take it; I don't think it
gets THAT much faster. But it is noticeably faster.
- Leo Butler of Brocade Communications
The design I'm working with has about 1.5M gates of standard cells. It
is a low-power SoC with complex clocking which means lots of timing
exceptions. I been using PT and PT-SI with this design about one year
now updating version by version from 2003.03 to 2004.06.
So to answer your benchmark question, I run update_timing with versions
V-2003.03-SP2 and V-2004.06, both on 32-bit Linux. With V-2003.03-SP2
the runtime is around 10 minutes and the memory usage 1.6 GB. With
V-2004.06 the runtime is around 5 minutes and the memory usage down to
1.2GB. So it seems that in one year Synopsys has been able to make PT
two times faster, which is great. But the bigger leap for me was when
last summer I was finally able to change from Solaris platform to Linux.
The update_timing runtime dropped from 30 min to 10 min. All this
together has made my PT environment 6 times faster in 1 year! This has
a huge effect on my output, especially when developing the constraints.
Also I've been able to cut down on my coffee consumption significantly.
I must say that the support from Synopsys is nowadays excellent when
comparing it to few years ago. At that time we had to give a large
pile of dollars to Synopsys to get help and usually all we got was one
guy on our site for one week complaining about our design style ("too
complex clocking, does not fit our methodology, this should be
re-designed..."). Nowadays they are actively calling me and asking
how to fit their tool to our design. So some competition does good for
all of us.
Then there is the sad story of OCV and CRPR. I have not used my old
faithful calculator this much since my university days. I've used hours
after hours checking and comparing CRPRs in different versions of PT.
If somebody has the time, it would be interesting to know the amount of
STARs related to CRPR in versions from 2003.03 to 2004.06. This has
lowered my trust on PT. Earlier it was the PT that was used to check
the results of other tools. Now I'm using other tools to check the
PrimeTime results.
- Pasi Tukiainen of Nokia
We use PrimeTime. 3X seems to be more improvement than we have seen
with vanilla PrimeTime. (1.5X seems more right.)
- [ An Anon Engineer ]
We use PrimeTime. Have no speed info.
- Larry Tesdall of QLogic Corp.
We have found that PrimeTime is faster if one uses the ILM feature.
- David Fong of S3 Graphics
We have seen some speed improvements in PrimeTime. From 1 hour down to
20 minutes on a 800 Kgates design. So it is true.
- Massimo Scipioni of STmicroelectronics
Certainly not! When you work around a bug in 2003.12-SP1 that forces us
to use "link_design -keep" (else we can't have design dbs in our
link_path), and when we set timing_report_unconstrained_paths to true
(otherwise run the risk of breaking custom timing reports) our speed up
factor goes to about 1.1X.
PrimeTime 2004.06 fixes our bug but a more intelligent behaviour is
needed for timing_report_unconstrained_paths. Simply, the unconstrained
paths should be removed from report_timing if there are constrained
paths, but, if I do a: "report_timing -to output_port" then I should see
the path, constrained or no. A little bit of smarts would give us the
speed gains without the usage pitfall.
- Andrew Bell of PMC-Sierra, Inc.
We use PrimeTime for STA. We didn't pay attention in the run time
lately. Our design is small so PrimeTime can always complete the
task over night.
- Edmond Tam of Global Locate, Inc.
We still use PrimeTime for timing closure. It has good capacity.
Maybe our chip is not large enough.
- Haiming Jin of Intel
Only use PrimeTime; it is faster than say 3 years ago, but not 3X.
Probably a unique corner case where they focused some attention on
that analysis type... Reminds me of the FPGA PREP benchmarks back
in the early 90's to try to avoid these unsubstantiated time
and QoR trials.
- [ An Anon Engineer ]
Don't use Prolific ProTiming or Sequence ShowTime. We haven't hit
any show stoppers that really necessitated us looking outside of
PrimeTime yet.
- [ An Anon Engineer ]
The most interesting data point I can give you is for PrimeTime. The
3X PrimeTime speedup is completly real. I see at least this much.
I met the Prolific Protiming guys at SNUG, bought the tool, and am a
happy customer.
- David Oliver of AMD
We use PrimeTime 2003.12. We have not seen much difference in runtime
vs. older versions. Our limiting analysis is post-layout with back-
annotated parasitics.
- Juan Carlos Diaz of Agere
Just PrimeTime used. We are stuck with old versions qualified for
production silicon. Newer versions are used and are faster. Again it
seems hardware improvements are the main delta in runtime improvement;
32-bit Linux or 64-bit Opteron.
- [ An Anon Engineer ]
We use the timing engine within DC for the most part. On the physical
side we rely on the timing engines within SoC Encounter, but doublecheck
against PrimeTime for sanity. I'm not sure about speed comparisons, and
we don't use any third party timing tie-ons. I think they complicate
the matter.
- Gord Allan of Carleton University (Canada)
Yes, we use PrimeTime and PrimeTime-SI. I'm a little down-rev, on
2003.03-SP1, because I'm only just finishing up a project. We try not
to change revs in mid-stream, since something usually breaks. PrimeTime
has been getting faster, but I got the best improvement by running
under IPF Linux.
No, we don't use any third-party tools with PrimeTime.
- [ An Anon Engineer ]
I always use PrimeTime and don't feel so improved speed.
- [ An Anon Engineer ]
We use PrimeTime, but we haven't benchmarked it.
- [ An Anon Engineer ]
PrimeTime performance has improved but I don't know if I can justify 5X.
Have looked at Prolific and it produces promising results on setup but
is lacking in hold capabilities.
- [ An Anon Engineer ]
PrimeTime has increased in speed. Do not use third party add-ons.
- John Blessing of Harris RF
We are 100% PrimeTime with no add-ons. CeltIC for SI though.
- [ An Anon Engineer ]
PrimeTime is our standard. Haven't noticed any speed-ups, though.
- [ An Anon Engineer ]
We use PrimeTime. Never noticed the 3X improvments, though.
- [ An Anon Engineer ]
PrimeTime is 3X Faster than what???? PrimeTime is now a bit faster but
not that much. It would be nice if Aart could either mention the
customers who had such improvement figures, or admit it was just some
internal test case.
- Marcello Vena of Xignal Technologies AG
Mmm... We have PrimeTime but are not using it because our design
(all single edge-trigger design) does not need it in pre-P&R phase,
and ASIC vendor does our post-layout analysis.
- [ An Anon Engineer ]
I have tried PrimeTime 2003.12. I don't have numbers, but it did feel
faster. My first problem is that they are introducing performance
improvements that change PrimeTime's behaviour (e.g. reporting
"no paths" instead of "path is not constrained"!). So I would not call
this a real improvement. However, I have seen one thing that we have
been asking for for years: a timing path table. It's still not quite
up there yet, but the GUI is finally worth using.
I asked for a few improvements. I don't know if they have been filed as
enhancements yet:
1) Sorting with respect to several columns (if the first field is the
same, go on to the next field for comparison). Right now, they can
only sort with respect to one.
2) Ability to save sorting preferences.
3) Enhanced filtering at the GUI (combining filter rules into complex
expressions if need be). It is not documented, but if you
right-click on the table, they give you filters to play with
already. :-)
4) Ability to save filtering preferences.
In short, PrimeTime is improving. We just disagree on what the most
important (and real) improvement is.
ProTiming: I'm not a user.
ShowTime/CoolTime: I'm not a user.
- William Natter of Nortel Networks
|
|
