( DAC'17 Item 2 ) ------------------------------------------------- [11/17/17]
Subject: Solido ML, BDA Ravi, Tom Beckley SSS makes #2 for Best of 2017
SOLIDO KICKS ASS: In a normal year, Solido Amit gives his usual variation
talking points along with his very widely read annual SPICE use survey.
263 engineers surveyed on their Custom IC design types & nodes
263 engineers on their present day SPICE use and SPICE leaders
And the variation part of Amit's 263 engineer SPICE survey...
But this year Amit added a little side pitch about lib characterization...
Amit added 263 engineers on Library Characterization this year!
... which he then later explained at the DAC'17 in Austin was really Solido
using Machine Learning for library characterization -- which got the most
interest for Custom IC in this year's "Best of 2017" EDA tool user survey.
(See the user comments below.)
---- ---- ---- ---- ---- ---- ----
YAY! RAVI DIDN'T LEAVE!: And normally in the SPICE & SPICE tools catagory,
I'd get a bunch of Mentor BDA AFS fanboys chatting up BDA AFS.
But this year, instead, they where doing a collective sigh of relief that
Ravi wasn't leaving the new "Mentor, a Siemens business" -- meaning BDA AFS
would still be supported and improving. (See user comments below.)
---- ---- ---- ---- ---- ---- ----
TOM BECKLEY IS UP TO SOMETHING: And in the SPICE Monte Carlo wars where it's
a tech food fight between Solido vs. Cadence TeamADE vs. MunEDA...
... something happened mid-2017 where Cadence did a quiet beta-launch of
their new proprietary Scaled-Sigma Sampling (SSS) approach to Monte Carlo
which popped up on my user survey this year. (And again, see the comments.)
"Where there's smoke, there's fire."
- American proverb
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
SOLIDO MACHINE LEARNING
Not only did Solido add machine learning-based algorithms into
their tools before Cadence and MunEDA, but also before many other
AI tools and services companies out there today.
They are definitely ahead of everyone else in the practical use
of the "Learning" part of ML. It requires a mountain of data I
believe they've already acquired over many years.
Side note: Solido also has a great support team.
---- ---- ---- ---- ---- ---- ----
I'm using Solido's Machine Learning Characterization tool.
Here's how we use it to identify anchor corners for characterization.
1. Characterize all corners for a sample of cells.
2. Use full characterizations to define the anchor corners.
3. Then fully characterize these anchor corners for all libs.
4. Solido ML Characterization then uses these corners to
predict the rest of the in between values.
We've validated this methodology for our std cell libs at 14 and 10nm
to prove it produces a valid data with very good accuracy.
We now use it to characterize everything -- delay, slew, CCS data,
power, leakage etc. ... - everything in our Liberty format.
---- ---- ---- ---- ---- ---- ----
Solido Variation Designer
Solido's statistical PVT tool analyzes the worst case SPVT corner
4-sigma designs with only 300 SPICE simulations.
It uses machine learning techniques to create a model and makes a
prediction. Then you run an iteration of simulations to tune/learn
and confirm the accuracy of the prediction.
You gain even more efficiency as the Solido tool continues to learn
more over time.
---- ---- ---- ---- ---- ---- ----
Solido ML characterization has generated some interest within my team.
It can save time by allowing PD engineers to get started on their work
while the actual library characterization goes on in parallel.
---- ---- ---- ---- ---- ---- ----
We are evaluating Solido's Machine Learning Characterization tools.
If Solido can predict a lot more corners in less time, we can spend
that time on more designs.
Plus we can use fewer cores as we try out different libraries.
---- ---- ---- ---- ---- ---- ----
Solido Machine Learning Characterization is a newer tool that uses
machine learning to do corner predictions with high accuracy. It
takes a set of basis corners, and can then fill out your corner
portfolio.
This saves days and days of SPICE simulation time.
---- ---- ---- ---- ---- ---- ----
Our company has a close working relationship with Solido. Their
products look good overall.
To me, Solido's most interesting technology is their Machine Learning
Characterization suite. It takes initial corners and uses the
information to predict new corners.
I want more details on the ML algorithms they use and how they do
their predictions, to confirm if they are robust enough.
---- ---- ---- ---- ---- ---- ----
Solido's variation analysis gives you an early estimation of what needs
to happen in your design to improve the process variation impact; which
is important to us for yield concerns.
We like working with Solido because their R&D likes listening directly
to customers. For example, they's adding machine learning to ID risk
areas in silicon. We don't see anyone else doing this.
---- ---- ---- ---- ---- ---- ----
Saw Solido' machine learning DAC panel -- their ML tools seem like
an innovative idea and a good change moving forward.
Something new and different.
---- ---- ---- ---- ---- ---- ----
Solido's machine-learning tools.
They significantly speed up the characterization and variation
analysis.
---- ---- ---- ---- ---- ---- ----
We've used Solido Hierarchical Monte Carlo to design 3 to 4 sigma
control logic, 4 to 5 sigma sense amps, and 6 to 7 sigma bitcells.
HSMC lets you verify your entire circuit without overdesigning,
i.e. rather than having to verify an entire critical path to the
6-7 sigma requirement for bitcells.
---- ---- ---- ---- ---- ---- ----
Solido does a good job with their statistical circuit analyzer.
---- ---- ---- ---- ---- ---- ----
Solido Variation Designer
- Minimizes failures due to variations in PVT
- Gives me a better understanding of designs with simulation
data, plus their analysis is good
I also look at Solido's ML tools:
- minimize simulation time by ML data mining of history;
very good approach.
- Data visualization and great user interface in the tool
suite
---- ---- ---- ---- ---- ---- ----
I was interested to know how machine learning is used in different
platforms, and Solido's machine learning session was enlightening.
We are trying to use ML for low power and security.
---- ---- ---- ---- ---- ---- ----
We used Solido High Sigma Monte Carlo (HSMC) for design signoff for
bitcells, sense amps, and critical circuits.
It has two modes, and I've used both of them:
- Regression mode. Solido uses a form of regularized linear
regression for when the output variable is continuous (e.g.
bitcell IDS, sense amp differential).
- Classification mode -- to identify if there's failure boundary
such as a binary pass/fail cliff
HSMC is absolutely required for design signoff.
---- ---- ---- ---- ---- ---- ----
Machine learning has coming a long way. It is model fitting -- think of
it as AI, such as modeling driverless cars.
Solido's ML speeds things up. They fit the curves on one set of data,
then the data set changes, and they characterize a new model.
There is supervised learning, where you have a problem and know the
answer, then fit it to a model. So, machine learning seems like
artificial intelligence, and it passes a Turing test. ('How are you?'
'I am fine.')
Even so, it doesn't work miracles -- such as vector generation for
PDN profiles for worst case currents. How do you do machine learning
on a problem that humans haven't yet figured out how to solve?
I'm waiting for the next generation of ML, which will have more insight
rather than curve fitting.
---- ---- ---- ---- ---- ---- ----
Solido's machine learning focuses on variation and lib characterization.
I understand it's a working solution and methodology.
I was looking for a way that machine learning could be used for digital
design verification, but if you do library characterization, Solido
would be worth it.
---- ---- ---- ---- ---- ---- ----
Solido discussed machine learning at their DAC panel -- in particular,
they discussed challenges with reducing sample counts to reduce total
simulation time. They did not make a strong argument for how the total
number of samples were reduced compared to Monte-Carlo sampling; in a
machine learning context, I would have expected a different discussion
and focus.
One thing Solido mentioned was an integration with modern tool chains
that supported the data extraction required for any kind of statistical
analysis. Given that this instrumentation is often most of the work,
this is extremely valuable.
I suspect many customers are much more interested in the abstraction and
support for data extraction and batch execution. I have encouraged data
oriented customers to look at Solido for this reason.
---- ---- ---- ---- ---- ---- ----
We are interested in looking at Solido Variation Designer.
---- ---- ---- ---- ---- ---- ----
I like the ML stuff Jeff is injecting into Solido tools.
---- ---- ---- ---- ---- ---- ----
Solido looks very promising for variance handling. We're banking on
their approach.
---- ---- ---- ---- ---- ---- ----
Solido Variation Designer. The first (and potentially only one for a
long time to come) concrete use of machine learning in EDA.
---- ---- ---- ---- ---- ---- ----
I didn't have the heart to tell Jeff that Calibre has been doing
machine learning for at least 5 years now.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
MENTOR BDA AFS & CADENCE SPECTRE
Having our primary SPICE (BDA) go under new mngmt was a scare.
Ravi's visit helped calm those concerns.
---- ---- ---- ---- ---- ---- ----
We're happy to see that Ravi and BDA are still at Mentor after the
Siemens buyout.
---- ---- ---- ---- ---- ---- ----
My company has invested million into BDA AFS.
Ravi leaving would have been disastrous for us.
---- ---- ---- ---- ---- ---- ----
Mentor/Ravi staying at Siemens was our top news for the year.
---- ---- ---- ---- ---- ---- ----
We went oh shit on the Siemens-Mentor acquisition. Ravi had to
fly in to convince my CAD group that nothing was going to change.
---- ---- ---- ---- ---- ---- ----
We're a big AFS house.
Ravi remaining at Siemens was important for us.
---- ---- ---- ---- ---- ---- ----
We're a volume user of BDA AFS.
---- ---- ---- ---- ---- ---- ----
AFS has the best support for our memory cell circuits.
---- ---- ---- ---- ---- ---- ----
BDA AFS and Cadence Spectre in ADE. Same as the last 3 years.
---- ---- ---- ---- ---- ---- ----
Mentor AFS
---- ---- ---- ---- ---- ---- ----
For high accuracy, we do AFS.
For price, we do Spectre.
---- ---- ---- ---- ---- ---- ----
Cadence Spectre-APS and Berkeley AFS.
---- ---- ---- ---- ---- ---- ----
Cadence Altos and Cadence Spectre-XPS.
---- ---- ---- ---- ---- ---- ----
Magma FineSim
---- ---- ---- ---- ---- ---- ----
FineSim.
---- ---- ---- ---- ---- ---- ----
Our SNPS salesman keeps telling us that we're the last remaining
original Nassda HSIM customer in the world.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
CADENCE SSS
You might want to look at Cadence SSS.
---- ---- ---- ---- ---- ---- ----
Cadence SSS
---- ---- ---- ---- ---- ---- ----
CSNS SSS. They fixed their WCD problems.
---- ---- ---- ---- ---- ---- ----
I don't think it's launched yet, but look at SSS if you can.
---- ---- ---- ---- ---- ---- ----
Virtuoso ADE-XL with SSS. Don't believe those Solido lies.
---- ---- ---- ---- ---- ---- ----
Cadence ADE-XL with Spectre
---- ---- ---- ---- ---- ---- ----
SSS for MC makes sense
---- ---- ---- ---- ---- ---- ----
Tom has SSS this year. You should check it out.
---- ---- ---- ---- ---- ---- ----
I don't know if SSS is public yet. Look into it when it is.
---- ---- ---- ---- ---- ---- ----
Beckley does a pretty convincing SSS talk these days.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
MUNEDA WICKED
We've known the MunEDA engineers for more than a decade now.
WiCkeD does the job.
---- ---- ---- ---- ---- ---- ----
MunEDA Wicked
resizing FEA, DNO, GNC, YOP
reliability DNO, CED, WCO, BAS, FEA
---- ---- ---- ---- ---- ---- ----
Don't trust the Solido HSMC for memories.
Do MunEDA instead.
---- ---- ---- ---- ---- ---- ----
By German law, the only SPICE tools we can use are MunEDA.
---- ---- ---- ---- ---- ---- ----
Related Articles
Solido ML, BDA Ravi, Tom Beckley SSS makes #2 for Best of 2017
Users say MENT Calibre is still DRC king was #3 in Best of 2017
Genus RTL synthesis gaining traction vs. DC is #4 of Best of 2017
IC Manage & Cliosoft get #5 Best 2017; MethodICs & Dassault missing
Join
Index
Next->Item
|
|
