( ESNUG 573 Item 4 ) -------------------------------------------- [06/09/17]
Subject: Amit added 263 engineers on Library Characterization this year!
Here's my third annual pre-DAC external survey on sponsored by Solido.
This year we have 3 years of data to do even better analytics on trends.
Our prior custom IC design surveys were: 2016 and 2015.
This year 263 designers and engineering managers responded to our 2017
survey. Below is what they reported about their general SPICE use and
SPICE requirements.
- 263 engineers surveyed on their Custom IC design types & nodes
From: [ Amit Gupta of Solido Design]
Hi, John,
We added Library Characterization as a new topic in our 263 engineer survey
this year. It's a ~$100 million market -- library characterization tools
drive SPICE simulator use, plus the .libs produced are used in static timing
analysis tools.
Feeds an Ecosystem: characterization tools are what create the .lib models
of standard cell, memory, and analog/IO design building blocks. These .lib
files are also used as input for timing analysis tools (e.g. Synopsys
PrimeTime and Cadence Tempus). Characterization is near the base of every
EDA chip design food chain pyramid.
---- ---- ---- ---- ---- ---- ----
ACCURACY IS THE BIGGEST HEADACHE FOR CHARACTERIZATION
Half of designers saw characterization accuracy as their #1 issue. They use
their design's .lib statistical variance to improve accuracy -- in the form
of AOCV (advanced on-chip variation), POCV (parametric on-chip variation),
or LVF (Liberty variance format) data.
Engineering man-hours, tight project deadlines, overdesign/over-margining,
and simulation/CPU constraints are second-tier issues -- but all are still
relatively significant.
In the survey finding below, we again see accuracy issues rise to the top.
This is consistent with the large number of designs being done sub-28 nm.
Getting accurately characterized libraries is a concern. With the increased
number of variants-per-process node (e.g. low power, high performance), the
amount of characterization needed is going up.
---- ---- ---- ---- ---- ---- ----
LIBRARY BLOCKS BEING CHARACTERIZED
Most people know standard cell and memory blocks are characterized -- the
fact that 62 percent are characterizing analog I/O blocks was a surprise.
The 34% memory characterization is done by teams developing their own memory
blocks; the rest either do not incorporate embedded memory in their designs,
or they use third party memory IP.
---- ---- ---- ---- ---- ---- ----
THE LIBRARY CHARACTERIZATION LEADERS
Cadence has the #1 technology leadership perception, with 62%. Synopsys is
a distant #2, with 26% naming them. Mentor ranked #3, with 6%.
Cadence acquired Altos in 2011 for its Liberate library characterization
tools -- with that acquisition, they are dominating the market. Synopsys
acquired Magma in 2012 in part for SiliconSmart, but the data says it
appears to not be doing all that well.
(I would have thought that Synopsys would be ahead in characterization, as
historically, most of library SPICE simulation was being done by Synopsys
HSPICE, FineSim and CustomSim.)
What's interesting is the amount of library simulation that is now being
done in Cadence tools also coincides with Cadence's growth in SPICE
leadership -- Cadence overtook Synopsys in memory SPICE (45%), matched
Synopsys in standard cell SPICE (45%), and also extended the Cadence
domination of AMS/Custom Digital/RF SPICE (67%).
Synergy: SPICE drives library characterization and library characterization
drives SPICE. They feed each other.
---- ---- ---- ---- ---- ---- ----
SOLIDO MACHINE LEARNING (ML) CHARACTERIZATION SUITE
Solido launched its Machine Learning (ML) Characterization Suite in April
2017. Our tool uses many of the machine learning techniques that we had
already developed and deployed in our Solido Variation Designer tools
since 2005.
Our ML Characterization Suite has two tools:
1. ML Predictor instantly and accurately generates new Liberty
models at new corner conditions. Normally you need to create a
.lib for every process/voltage/temp condition (e.g. FF, 1.2V, -40C),
Vt family, supply, channel length, model rev -- and it typically
takes a ton of time to do a characterization run.
ML Predictor models the Liberty space using sparse data from
existing Liberty models at other corner conditions. Let's say
you have 50 corners you need .libs for. You characterize your
first 26 corners. Then use ML Predictor to characterize the
remaining 24 corners instantly -- without using more characterization
or SPICE simulation licenses -- all because of the ML model
that ML Predictor had built from the earlier characterized training
data. ML Predictor doesn't use SPICE licenses, but it does take
advantage of a CPU cluster.
By using machine learning, ML Predictor reduces your library
characterization time by 30% to 70%.
2. ML Statistical Characterizer delivers true 3-sigma statistical
timing data (LVF/AOCV/POCV) values with Monte Carlo and SPICE
accuracy, including non-Gaussian distributions. It adaptively
selects simulations to meet accuracy requirements and to minimize
runtime for all cells, corners, arcs, and slew-load combinations.
The industry has added 3 statistical .lib formats over the years.
It started with AOCV. Then Extreme-DA created the POCV standard,
and the company was acquired by Synopsys. More recently, the IEEE
Liberty Technical Advisory Board (LTAB) standards body came out
with the LVF format.
The driving reason for putting in statistical information into the
.lib is to get rid of under- and over- design, which is even more
of a headache at sub-16nm nodes and for low power design, where
you have less margin. The issue for the designer is that adding
LVF/AOCV/POCV data into the .lib blows up the required simulation.
Solido Statistical Characterizer gives Monte Carlo accuracy in
1000X fewer simulations by using machine learning.
---- ---- ---- ---- ---- ---- ----
LIBRARY CHARACTERIZATION IS CHANGING
Library characterization is important to core EDA tools. The .libs are
created by SPICE simulators and feeds into static timing analysis. The
major headache in library characterization is accuracy.
This year machine learning is already improving lib characterization
accuracy and run time.
- Amit Gupta
Solido DA San Jose, CA
---- ---- ---- ---- ---- ---- ----
Related Articles:
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...
Amit added 263 engineers on Library Characterization this year!
Join
Index
Next->Item
|
|
