Subject: Rivals laugh at Synopsys annual PrimeTime 2X speed-up claims

> MOUNTAIN VIEW, Calif., Jan. 11 -- Synopsys, Inc. (NASDAQ: SNPS), a world
> leader in blah blah blah today announced the immediate availability of
> PrimeTime 2009.12, delivering up to 2X speed up of timing signoff through
> the addition of threaded multicore processing.


From: Dave Desharnais <dez=user domain=cadence not calm>

Hi, John,

When I first saw this press release I asked myself: "Why is Synopsys even
announcing this?"  Then I remembered it's a Synopsys tradition to begin each
new year claiming a 2X speed-up for PrimeTime.  Their mistake this year is
that it's the SECOND time they've added "multicore processing" to PT.  Oops!

On Jan 11, 2010, SNPS announced for PrimeTime 2009.12:

  "Synopsys today announced immediate availability of PrimeTime 2009.12,
   delivering up to 2X speed up of timing signoff through the addition
   of threaded multicore processing."

On Feb 12, 2009, SNPS announced for PrimeTime 2008.12:

  "The latest release, PrimeTime 2008.12, includes a flexible multicore
   processing technology that makes more effective use of both single-core
   and multicore CPUs across today's compute server farms, harnessing
   their compute potential.  ...to deliver up to 2X faster runtime..."

On Jan 23, 2008, SNPS announced for PrimeTime 2007.12:

  "Broad improvements in design data reading and linking, intelligent disk
   caching, incremental timing updates and fine tuning of algorithms have
   resulted in an average 2X runtime improvement and 33 percent memory
   reduction over PT 2006.12 while maintaining golden signoff accuracy."

2X on 4 CPU is not a breakthrough.  We've been doing this for nearly 2 years
already with Encounter Timing System (ETS).  However, we have a very
different approach to how we accomplish this, plus our new memory
architecture gives us a better starting point, even on a single CPU, where
we are consistently 40%-60% faster in benchmarks versus PrimeTime.

What's more, reading between the lines, the way that Synopsys is doing this
is not actually true multithreading in that it is not shared memory space.
The caution is although the performance of PrimeTime may be faster over
4 CPUS, the memory/capacity can take a hit.  Gotta have a slick compute
farm at your disposal, or good IT folks on hand.  That can be a problem.
Designs are not shrinking last time I checked, and with more complexity
due to managing process variation and multi scenario analysis (MMMC) comes
more memory demand.

Also, since this is "signoff", accuracy should be paramount for a customer
signing off a design, breaking up a design and sending it across a network,
opens up the question of accuracy.  For example, how do you model SI impacts
across the boundaries of these partitions?  A separate machine is a separate
machine.  So, the question is what are customers having to trade off for
this gain in PrimeTime performance?  And is this something they really want
to have to manage?  Signoff is the final frontier.  Getting to the wrong
answer faster is counter to the entire concept of signoff.  You want speed,
capacity, AND accuracy, ALL at the same time.  At least that is how we
approach it here at Cadence.

What we've seen over the past 2 years when we engage with customers using a
PrimeTime signoff is a churn at the end of a design flow as they struggle to
converge on timing, SI, power, and the interdependencies.

MMMC timing, SI, power, etc. has to be in concert; not sequentially - AND it
has to be right there inside the physical implementation domain as well.

  a) as a pure analysis step, you need to go beyond the traditional
     timing/SI analysis, and consider IR drop impacts on delay, jitter,
     thermal impacts, DFM, etc. concurrently and
  b) handle this in implementation so you reduce churn and iterations at
     the end of the project.

This is something we uniquely do at Cadence w/ our Encounter Timing System,
Encounter Power System, and Encounter Digital Implementation System.  This
approach resonates with designers, and that's what we're after.

    - Dave Desharnais
      Cadence Design Systems, Inc.               San Jose, CA

         ----    ----    ----    ----    ----    ----   ----

From: Isadore Katz <is=user domain=clkda not calm>

Hi John,

It's Isadore.

Concerning the recent PrimeTime multi-threading announcement, I say to SNPS
"welcome to the threaded timing club!"  (I thought they had announced this
same multi-threading claim last year, but I guess anything worth announcing
once is worth announcing twice.)

I do not know "2X compared to what", but assuming it is a legit increase,
the PT team should be congratulated.  Making 15 year old code go faster is
never easy.

Having said that, 2X on four processors is not that impressive.  The
partitioning approach used by Synopsys does not scale well, and is not going
to address the other major issue for PrimeTime, memory footprint.  While
this might meet the immediate needs of some Synopsys customers, it does not
legitimately deal with the long term viability of PrimeTime architecturally.

From a performance & memory standpoint, our view is that the real technical
leadership is coming from CLK-DA (us) and Extreme DA.  Extreme on memory,
CLK on performance.  If it weren't for accumulated undocumented behaviors in
PrimeTime that have come to be called "sign-off", either of our tools
represent a much better and accurate technical path -- and in all likelihood
higher quality silicon.

For example, our Amber scales nearly linearly on 12 processors, and gets
continuous gains all the way through 16 processors.  We have also threaded
design reads and reporting to get a 5X to 10X throughput advantage over the
current release of PT.

We think our architectural underpinning is critical for two reasons. 

 - First, when one starts to work with things like statistical, the
   calculations become even more strenuous.  You need to leverage all
   of the processors available.

 - Second, 2X on four processors points to the limitations of PrimeTime
   over the long haul.  If this is the best that Synopsys can produce
   after several years of hard work, it points to the need for a
   bottoms-up re-write of the product.  This may well be underway (SNPS
   doesn't keep me current on this stuff) but some times you just need
   to start from scratch to get it right.

It all boils down to making 15 year old PrimeTime code go faster ain't easy.

    - Isadore Katz
      CLK Design Automation, Inc.                Littleton, MA

         ----    ----    ----    ----    ----    ----   ----

From: Arthur Wei <arthur=user domain=incentia not calm>

Hi, John,

Did you see that Synopsys PrimeTime press release today?

I think Synopsys is doing some catch-up here!  Incentia TimeCraft has had
multi-thread capability for multi-core CPU for two years now.

Their 2X speedup is not so impressive.  Our latest results show TimeCraft
can easily achieve 2X to 5X speedup in report timing.  TimeCraft also
supports multi-mode multi-corner (MMMC) timing analysis by submitting
jobs to network machines, and running in parallel under multi-thread to
maximize the machine performance. 

We also guarantee our results from multi-thread and single-thread to be
exactly the same.  TimeCraft's multi-thread technologies are much more
advanced.  We cannot wait for the next press release from PrimeTime, it
will probably be on location-based on-chip variation (LOCV) which Incentia
customers have been using for years.

    - Arthur Wei
      Incentia Design Systems, Inc.              Santa Clara, CA

         ----    ----    ----    ----    ----    ----   ----

  Editor's Note: I phoned and emailed multiple people multiple times
  at Extreme DA to see if they wanted to chime in, but nobody there
  got back to me.  The last update on their web site was from DAC on
  July 2009.  Anyone know if Extreme DA is still in business?  - John