( ESNUG 499 Item 1 ) -------------------------------------------- [02/23/12]
Subject: Users voting to save Magma Tekton and QCP
SPEED, SPEED, SPEED: From what I see of these user comments (and benchmarks)
below, users are seeing 5-10x faster runtimes for Tekton vs. PrimeTime and
it correlates with signoff. The other takeaway was that Magma's QCP just
leapfrogged Magma's Quartz RC by up to 10x also.
"Assuming the Synopsys-Magma merger goes through, as a Magma user,
which specific Magma tools do you want to survive and why?"
---- ---- ---- ---- ---- ---- ----
We have used Magma's Tekton tool now for over a year and recently just
started using Magma's QCP tool.
Both tools have proved to be very beneficial when it comes to turnaround
time and hitting our aggressive schedules. We have been using Synopsys
PrimeTime since 2000 as our official signoff timer, but just recently
have forecasted moving to Tekton as the official signoff timer in June
of this year.
The reason, runtime. Tekton has a decrease in runtime of 90% over
PrimeTime!
Million placeable instance design with 24 scenarios
Same number of CPU's and licenses (1 Primetime license versus 1 Tekton
MMMC license, each utilizing 4 - 2.6GHz Linux CPU's)
Tool Runtime
--------- ----------
Tekton 1.9 hours
PrimeTime 18.5 hours
Tekton's accuracy is within pico-seconds of PrimeTime, usually on the
pessimistic side.
This is another metric that is necessary when moving to a different
signoff timer. We recently have been using QCP as our digital
extraction tool because it more accurate to QuickCap (an industry
standard 3D capacitance extractor) than Magma's old extraction tool,
Quartz RC. We have seen a decrease of runtime using QCP by 91%
compared to Quartz RC! QCP can achieve this runtime by doing
multi-corner extraction concurrently on multiple CPU's.
Tool Runtime
--------- ----------
QCP 1.9 hours
Quartz RC 22.2 hours
Combining QCP and Tekton in our tool suite saves us roughly 37 hours in
turnaround time for this size block. I want to stress that this is 37
hours of time saved for *every* ECO performed on this design.
Flow Runtime Savings for each ECO
--------------------- ---------- --------------------
Quartz RC + Primetime 40.7 hours
QCP + Tekton 3.8 hours 37 hours
This is a metric no other vendor can match. It would be detrimental to
the ASIC design community to abolish these tools and waste this valuable
IP just to rid competition.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
We started evaluating Tekton/QCP last year. Our main concern was its
correlation with PrimeTime, followed by performance improvements.
We used two test cases from 65 nm and 90 nm production designs with
NLDM/CCS libraries to evaluate how Tekton timing engine correlated to
our sign off tool. We saw a close comparison using zero wire load,
lumped C and coupled C for SI (with Lump C and coupled C used from sign
off SPEF).
For the first test case comparing a little over 10 K paths we got:
avg diff abs diff std dev
-------- -------- -------
ZWLM (ps) 0.11 2.06 3.11
LUMP C (ps) 9.98 13.08 13.53
SI (ps) 18.85 20.32 17.39
With respect to performance in the current PrimeTime flow we have to
extract to SPEF and go into our timing tool for multi-scenario analysis.
We are looking into eliminating this cumbersome and time consuming file
transfer with the Tekton/QCP built-in extraction/timing analysis.
In our preliminary evaluation, done with a typical ARM-based design with
5 modes and 4 extraction corners, we saw a combined (timer + extraction)
time as follows:
PrimeTime sign-off 1 hour
Tekton/QCP 11 minutes
We see great potential for performance improvements here as our designs
keep increasing in number of modes as well as extraction corners.
We would miss a great opportunity to improve the performance and QoR of
our existing design flow if Tekton/QCP were to be discontinued after
Magma acquisition by Synopsys.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
We have successfully deployed Tekton in our company to shorten our ECO
cycle and hence would like Tekton to be saved.
We found out Tekton is significantly fast, about 2X-3.5X faster compared
to our PrimeTime signoff tool, including SI and non-SI.
Tekton's correlation with PrimeTime was really good, over 99.5% of the
paths were within 2-3% for some of our blocks. This made Tekton our
preferred tool for backend timing debug as well fast what-if ECO
analysis.
Block Tekton PrimeTime Tekton PrimeTime
(minutes) Runtime Memory (MB) Memory/mode (MB)
(minutes) (6 modes in design)
------- --------- ------------ ----------- -------------------
block1 18.17 38.46 2262.7 1616.53
block2 95.32 224.28 4796 1729.68
block3 91.15 339.98 4767.4 NA
block4 141.22 374.72 6777.5 1782.21
block5 159.1 422.53 6377.4 1852.24
block6 89.33 247.57 5307.7 1801.28
block7 18.52 37.15 2247.2 1428.40
block8 125.16 415.31 5679.1 1739
block9 205.51 680.71 8366.6 1812.82
block10 75.38 189.89 4495.5 1690.81
block11 23.18 48.96 2680.7 1618.32
block12 268.56 532.84 11252.8 2231.67
block13 12.47 25.34 2262 1575.45
block14 91.2 191.26 4748.8 1708.13
block15 51.6 112.37 3519.8 1648.59
block16 5.42 7.25 1853.3 1569.45
The ability to do a MMMC analysis using one license was another
extremely useful feature of Tekton. We were able to develop scripts for
hold fixing by inserting buffers at the points where we had huge
positive margin in setup. This helped us in reducing the number of
iterations required for hold fixing.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
Tekton. PrimeTime is the industry accepted tool, so it's understandable
that this is a tough one. The simplicity and performance of the tool
are impressive though.
We recently converted our PT scripts over to Tekton scripts, it was as
simple as linking in the .lib files instead of the .db files. Runtimes
were 1/6th what we were seeing in PT and given the support for PT
commands it was a quick bring up with the timing guys.
- Mike Lafferty of Availink
---- ---- ---- ---- ---- ---- ----
We don't have Tekton ourselves, instead we relied on the UK Magma AEs to
run Tekton/SI for us. Since Tekton accepts Volcano, getting this up
and running was quite quick. On the timing side, we found that Talus is
more optimistic in some corners (hold timing, both fast and slow
corners), more pessimistic in corners (setup).
With a sub-50K instance design, running both Talus and Tekton is quick,
but I must say I was baffled when the UK AE at one point turned around
the database in less than an hour... at 9.30pm, on a Friday evening. So
he saw our email, downloaded our Volcano, ran Tekton in all corners
(setup, fast hold, slow hold, all with/without SI), and replied with his
results all in an hour. It might have helped that the mother-in-law was
visiting on the particular evening. ;-)
We also had Synopsys services org run PrimeTime-SI analysis in parallel
(as a safety measure) and getting this up and running took almost two
weeks just to get the LIB warnings (and errors) ironed out. It seems
that PT is much more picky about the LIB contents, much to our chagrin
(most of the IP is internally developed, and nobody volunteers to write
LIB files, so the LIB quality is less than stellar).
Once set up, we found Tekton's timing to be very close to PT-SI (<0.1%),
but Tekton and PT-SI not agreeing fully on the interpretation of the LIB
files (if you don't provide a logic function from input to output of a
cell, Tekton will say that there isn't a timing path, while PT-SI will
say there is a path).
We had a few paths where the hold timing between Tekton and PT-SI
disagreed by 400 psec, but this was tracked down to shoddy timing in a
LIB file (discontinuous timing in the hold timing matrix).
We're in the process of switching to a new technology node, and are
planning on using Tekton to ensure that the Talus SI/STA setup will
track the silicon for the first test chips.
- Kai Harrekilde-Petersen of GN Resound DK
---- ---- ---- ---- ---- ---- ----
Tekton should stay. In fact I strongly recommend Tekton's ground up MCMM
technology to be studied by PT R&D and learning incorporated.
We have been evaluating Tekton, working closely with their R&D/AE teams
on one of our 1.5 M 65 nm designs, evaluating across Synopsys PrimeTime
(and Extreme-DA).
- We had traditionally seen more than 12 hour runs on 20 cores
[3 machines] (DMSA) across ~100 scenarios with PrimeTime.
- This has been reduced to an impressive ~2 hours in 4 cores
(single machine) run in Tekton.
- That's a reduction from 240 CPU hours to 8 CPU hours, not to
mention from 20 licenses to a single license.
We see Tekton being very nimble and generally correlating very well with
Primetime & PT-SI. Initially during our eval, Tekton had limited
capability with SI-Noise using CCS-N libraries. In working with us
closely they have enhanced the tool and now we are seeing very good
correlation numbers across the two tools to around 99.97% across
different design scenarios.
We also did a detailed feature investigation across ~150 requirements
and found Tekton meeting 99% of them, the ones they didn't meet they had
plans and ETA.
In summary, we want Tekton to stay.
- [ An Anon Engineer ]
---- ---- ---- ---- ---- ---- ----
We recently switched to Tekton from Magma Quartz Time. As a new Tekton
user I have been focused on validating the basics first and will try
some of the advanced functionality such as MMMC later.
We have already taped out a chip using Tekton as a sign-off tool.
- Tekton is faster than both Quartz Time and the competitive commercial
STA tool. Sorry, no data here.
- It was easy for us to migrate from the Quartz Time environment to
Tekton, in part because of our familiarity with using a competitive
STA tool.
- The tool quality is good, much better than Quartz Time, and Magma's
support has been excellent.
QCP is Magma's sign off extraction tool, with a 3D field solver inside
as algorithm. We switched to it from Quartz RC.
We typically extract 5 corners for each design. I have data for one of
our designs:
- Quartz RC took 8 hours to extract each corner. We extracted all 5
corners in parallel, so the elapsed time was also 8 hours.
- For the same design, QCP only took 30 minutes to extract all 5
corners. i.e. QCP took 30 minutes per corner, but it was just one
run since it does true concurrent multi-corner extraction.
We would like Tekton and QCP to both continue to be supported
post-acquisition.
- [ An Anon Engineer ]
Join
Index
Next->Item
|
|
