( ESNUG 538 Item 9 ) -------------------------------------------- [04/04/14]
Subject: Detailed user eval of Cadence ADE-XL, Solido, BDA AFS, Spectre
> Our automotive customers are very serious about accuracy. Without Solido
> HSMC, it would simply be impossible to accurately test the sensitivity of
> our sense amp. We actually use Solido HSMC results to convince our
> customers of the accuracy of our sense amps.
>
> - Kenji Noda of NSCore, Inc.
> http://www.deepchip.com/items/0535-04.html
From: [ Mr. Magoo ]
Hi, John,
Our design team works on high-speed analog and analog/mixed-signal designs,
currently at 40 nm and 28 nm CMOS. We make digital baseband optical and
networking components; high-speed I/O; and high-speed DDR4 memory systems.
Because of bad experiences we had with Cadence ADE-XL variation analysis:
- ADE-XL doesn't let us pick different distributions.
- Sensitivity analysis in ADE-XL 6.1 doesn't work as it produces
illogical answers. E.g. it told us that our offset variation
was due to a digital line that didn't even change during the
variation. That's crazy.
We evaled and bought Solido first for 3-sigma Monte Carlo and then later
for 4- and 6-sigma HSMC. So for variation analysis, we primarily use:
Solido with ADE-L and Spectre
or
Solido with ADE-L and BDA AFS
but NOT
ADE-XL with Spectre nor BDA AFS.
We use Solido iteratively during design; it requires only a fraction of
SPICE runs needed for 3-sigma design than with Spectre or BDA AFS alone.
Solido usually needs less than 100 simulation runs.
# simulations
Tool for 3-sigma Monte Carlo
------------------------------- -----------------------
Spectre or AFS alone 400
Solido Fast MC + Spectre or AFS ~50
We also tried BDA AFS's rev of "Fast MC", but it's primitive, so we still
prefer Solido.
---- ---- ---- ---- ---- ---- ----
STATISTICS PACKAGES
Solido Fast Monte Carlo - contains a well done statistics package. It has
great histograms, showing the outliers so you can tell how tight your design
is. It also gives me a distribution, and confident intervals in its GUI.
Cadence ADE-XL - stats package inside is less than helpful, e.g. it
doesn't even have good histograms.
Solido Fast Monte Carlo -- shows which simulation failed.
Cadence ADE-XL - I must do 100 simulations and it just say 1 simulation
fails. Trying to find which one failed is trying.
VARIATION ANALYSIS GUI
Cadence ADE-XL's GUI for variation analysis is really bad. Part of the
problem is that unless you've read the manual, you don't know how to open
sensitivity analysis or histograms (it turns out there are multiple ways).
Cadence keeps buying tools from different companies and bandaids them
together.
The initial display is horrible. ADE-XL displays every device and gives you
a color to indicate results, not a correlation coefficient. Color just
doesn't do it. Plus each device has a little tiny infinitesimal graph with
a slope and you must compare those to figure it out. You must then find the
pull-down to get correlation coefficient between 0 and 1 -- it's not the
default. Plus you need a Cadence ADE-GXL license to use it!
Solido has a user friendly environment for variation. Its GUI is intuitive
and has better graphics. I was a first time user, and skipped reading the
user manual. A window just pops up, I look at it. Every time there is a
question mark, and if you click on it, it tells you what it does. You can
just navigate it as you go.
VARIATION SENSITIVITY ANALYSIS
Cadence ADE-XL environment with Spectre does have some Monte Carlo results
reporting, but some of its reporting is crude. Also, it's slower because it
requires more simulation runs than Solido for 3-sigma (per numbers above).
Solido has better sensitivity analysis than Cadence. Solido DesignSense
sensitivity analysis utility is great. It indicates which devices heavily
impacts parametric variation. For example, it points to which devices have
contributed the most to variation in the specification parameters such as
offset. Such problems are not always intuitive. Then I figure out how to
FIX it. (That's why my company pays me the big bucks.)
---- ---- ---- ---- ---- ---- ----
4- TO 6-SIGMA MONTE CARLO ANALYSIS
The number of SPICE simulation runs for random Monte Carlo analysis goes up
exponentially with higher sigma design.
Stand-alone SPICE
Sigma Probability Monte Carlo Solido
of failure simulation runs simulation runs
--------- -------------------- ---------------- ---------------
3 sigma 1 failure in 1000 5,000 30-1500
4.2 sigma 1 failures in 75,000 500,000 3000-6000
6 sigma 1 failure in a billion 5 billion 3000-6000
There's pain in designing to a 6-sigma level -- if you have low volume
designs with a high enough margin, you don't do it. But if you have
very cheap jelly bean parts with no margin, e.g. $0.001 per part, you do
need 6 sigma.
I used Solido HSMC for 4.2 sigma on all 3 of my regulators and their loads,
with 400-500 transistors. Cadence has nothing that compares with it.
And we are running the latest version of ADE-XL.
I can tell you horror stories. Cadence ADE-XL uses Latin Hypercube and WCD.
I was trying to see if we had even a 3-sigma problem, i.e. 1 failure in
1000. I ran 100 simulations using Cadence ADE-XL with Spectre and got a
failure. I wasted DAYS on that one corner that Cadence identified, and it
turns out it only had a 1 in 7 billion chance of error. That's the lifetime
of the universe!
We can't afford to waste days -- our schedules are tight. ADE-XL is not a
good tool or a good use of my time.
(BDA doesn't have an HSMC equivalent either.)
---- ---- ---- ---- ---- ---- ----
SOLIDO FAST MC VS. HSMC
As to when we use Solido Fast Monte Carlo versus Solido HSMC? We use HSMC
for two key purposes:
- Part of the characterization process for designs more than
3-sigma, e.g. our 4.2-sigma designs.
- Once we have completed our final design, we often want to
do a higher sigma analysis, even for 3-sigma designs.
When we get to a sigma level where Solido Fast Monte Carlo asks for way too
many SPICE simulation runs, we go over to Solido HSMC.
---- ---- ---- ---- ---- ---- ----
PARALLEL SPICE RUNS
Solido collects data from many Monte Carlo runs. Solido also runs parallel
simulations, so you can also submit as many simulations as you have available
on your farm. I used Solido to run 1000 jobs with 300-400 simulators at a
shot with no problems.
If the SPICE licenses are already being used, then Solido queues up the jobs
to run as soon as the simulators are available.
Cadence ADE-XL also offers a parallel simulation split.
---- ---- ---- ---- ---- ---- ----
USE SPECTRE OR AFS?
Solido doesn't care what simulator we use, it works equally well with both.
So I switch back and forth between SPICE simulators. Cadence is dirt cheap,
but slower. I only use Spectre when I run out of BDA AFS licenses, because
AFS is 6X faster than Spectre.
The Solido/AFS combination is as fast as you can get.
Also, doing random Monte Carlo with BDA AFS and/or Cadence Spectre can give
us different answers. So we use Solido as the arbiter for variation -- they
understand the problem.
For example, an engineer in our group had been tracking down a weird Monte
Carlo effect in a charge pump for weeks. He finally ran Solido/Spectre.
He found out the problem is real, and he understands it -- versus it just
being a simulator anomaly.
---- ---- ---- ---- ---- ---- ----
WHERE SOLIDO NEEDS HELP
- Solido Variation Designer is well integrated with Cadence ADE-L,
but not with ADE-XL. We start out with our simulation set-up
in ADE-XL; we then create corner files in ADE and save them.
We must then recreate the corner in Solido. Our understanding
is that the lack of a deeper interface is a Cadence restriction.
However, we create an XML file, and Solido could still write an
interface to read it in.
Part of this is due to the fact that we have 3 different
environments. It wouldn't be a problem if we could stick with
one environment (i.e. ideally we would have an ADE-XL and Solido
integration), as we don't want to manually create 15-20 corners
in Solido, and then in ADE-XL. But Solido is not cheap.
So we must fight for the Solido licenses, and the "losers" must
run ADE-XL for variation. Solido is well-integrated with ADE-L
so the porting issue doesn't arise there.
- Currently, Solido does not work with Cadence AMS Designer. AMS
Designer is used for simulating analog blocks and Verilog digital
blocks together in ADE L/XL environment.
I think there will be more need for analog designers to run AMS mixed-signal
simulation in the future with Monte Carlo mismatch.
---- ---- ---- ---- ---- ---- ----
WE AVERTED AN EXPENSIVE FAILURE
We did Monte Carlo analysis with Cadence ADE-XL and Spectre on one of our
designs. It didn't uncover any errors. We were about to take this design
to production.
We then ran it through Solido/Spectre and found a start-up problem in a set
of our fast regulators -- which would have limited our yield to only 1000
parts per million. Several devices had mismatched!
After Solido HSMC identified the problem devices, we were able to tweak the
design to get failures down to 1 part per million.
This incident would have been very expensive. We would have had to try to
figure out in the lab why our chip didn't start up right. This experience
showed our MGMT that it was prudent to have Solido HSMC as part of our
final design verification sign-off.
- [ Mr. Magoo ]
Join
Index
Next->Item
|
|
