( ESNUG 588 Item 1 ) ---------------------------------------------- [09/12/19]

Subject: Joe Sawicki on ML, Calibre, Solido, VC funding, and heuristics

              DAC'19 Troublemakers Panel in Las Vegas, NV

   Cooley: Joe...  Sawicki...  Your Solido got #1 -- I actually had to
           write this out -- your Solido got #1 in the Best of 2018 for
           EDA tools with Machine Learning.  (see DAC'18 #1)

           And then recently in May 2019, you did another AI/ML toolkit
           for Calibre "to help complex machine learning IC architectures
           and adding an AI/ML infrastructure to Calibre Machine Learning
           OPC (mlOPC) and Calibre LFD." (see Mentor PR 05/23/2018

           Can you put that in English?

  Sawicki: Likely not?  Was that a question?  [laughter]

   Cooley: I mean, what, are you [Seimens/Mentor] doing?  Just slapping
           on Machine Learning everywhere, like back in the old days
           where everything was "Open Systems"?

  Sawicki: Absolutely.  Just put ML on everything and it takes care of
           all the problems.

           I mean basically, what's happening... I think I can find a
           question in there...

           As I was saying earlier, it's always fun when the guy who
           moderates a panel is actually the troll on your panel.
           [laughter]

   Cooley: Rrrr.

  Sawicki: Hey, fair?

   Cooley: Ok, all right, I get it.

  Sawicki: There are a couple things going on.  I think ML sits on both
           sides of the house -- both our customer base and on the EDA
           side of things.

           And I don't think Mentor's unique in this place.

           One, is that you've seen this explosion of funding coming back
           in from venture capital into the IC space, and it's virtually
           all around chips that are implementing machine learning.

           Either really big ones for data center applications, or really
           small ones that are moving towards edge applications.  

           All of us are trying to put together technology that enables us
           to make that ML stuff easier for those guys to design.

           So, Calypto is an example of being able to do architectural
           design at C level, then intensive flow applications, and then
           automatically synthesize those into RTL targets for either
           FPGA or ASIC.  

           So that's the IC side of the ML house.  There's a bunch of other
           ML stuff too in terms of being able to handle that massively
           distributed memory architecture, how you do test?, how you do
           verification?

           And then the EDA tool side, is that machine learning is a really
           really useful algorithmic basis to do really large improvements
           in EDA tool performance. 

           Whether it be from things around doing learning -- to around
           EDA tool setup -- to where I'm sure everybody's out to test
           place and route tools -- and putting together environments
           that allow us to optimize the place and route set up.  Or
           whether it is stuff around anything that's pattern-based
           like DRC/LVS.

   Cooley: Right.

  Sawicki: Where you can look at Solido being an example of that, and OPC
           is another example where you do this application on top of
           your normal application, and you can get 10X or a 100X type
           of performance improvement.

   Cooley: But what's the difference between machine learning, and just
           what they used to call heuristics?  It's the same thing,
           isn't it? 

  Sawicki: That's like saying that a tricycle is just like a Harley.
 
           Heuristics are: "I've noticed a couple of things, and I
           can list five of them."

           Machine Learning is: "I've gone off of a database of billions
           of instances.  I built a massive inference engine.   And I
           will not only notice the things that I directly saw but I will
           notice things that are imperceptibly the same to the human
           perspective and actually give outcomes for those."

           When we model, for example on the OPC, where we'll do machine 
           learning OPC, we will literally model billions and billions of 
           shape contexts, and then put those through a training module to 
           put together one of these large neural networks.  

           That's what we use to do the first set of transformation.  And
           we do that 10X faster than what you [a human] could do by
           actually doing a  simulation-based technique.  And I'm sure
           virtually everybody who's on this panel is doing some project
           like this.  

           I mean, I can't even count how many projects we have internally
           to apply machine learning towards EDA application.  

   Cooley: You actually step in and check to make sure it's doing something 
           intelligent vs. something weird? 

  Sawicki: Well, one, good luck...  But there are aspects where it depends
           on the type of application.  With OPC, we'll use machine
           learning to do 95% of the fix and then we wanted to run a
           traditional simulation loop, do the final check.  

           You basically have that same thing you've been doing for years.  
           It runs as a final check.  Some of the other stuff is more 
           statistical.

           But the ML stuff works and I guarantee you there's not an EDA
           company out there that's not running 10 to 20 ML projects at
           this point to put AI into their products.

   Cooley: Okay.

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

Related Articles:

    Joe Sawicki on ML, Calibre, Solido, VC funding, and heuristics
    Anirudh on each Machine Learning engineer is worth $10 million
    Mo Faisal on analog IP inside $50 billion worth of AI/ML chips

   !!!     "It's not a BUG,
  /o o\  /  it's a FEATURE!"
 (  >  )
  \ - / 
  _] [_     (jcooley 1991)