( ESNUG 586 Item 4 ) ---------------------------------------------- [04/18/19]
Subject: Dean, Anirudh, and Sawicki on Big Data versus Machine Learning
DAC'18 Troublemakers Panel in San Francisco
Cooley: Dean.
Dean: Yes.
Cooley: You had a product, it was a tape-out predictor product that was
like a hot baby, and... (see DAC'16 #8)
Dean: It's still hot.
Cooley: Well it had "Big Data" written all over it.
Dean: Yep.
Cooley: Now everything is Machine Learning, did you guys not get the
memo?
Dean: No, I missed that memo. No, we're still, we [IC Manage]
are the Big Data company, we're the Design Data company...
Cooley: But you're not Machine Learning.
Dean: No, we are not Machine Learning. We are not AI. It's not been
our focus, but if you want to do Machine Learning or AI you
actually need a lot of data to do the training.
And so our focus is still -- we introduced a new, Envision
Verification Analytics [used on the Xilinx Zynq chip -- see
ESNUG 0550 #5] which was focused on using Big Data to let people
track, understand, observe in real time; actually one of our
customers described to me it's like a live TV feed of what's
going on in verification. I thought that was a pretty cool
description.
But it's doing the predictive stuff, and the data is being used
to train some AI and some machine learning -- but that's not
what we're doing.
And if you look at the market, the Big Data market is about $150
billion dollars now and growing to $200 billion, and the AI
market is ~$12 billion growing to around $30 or $40 billion.
So I think we [IC Manage] are in the right market [Big Data] for
right now, and I'm sure there's will be some great applications
of machine learning and AI in the EDA biz -- but I think we are,
as Joe Costello would say -- very very early and it's really kind
of unclear as to what is going to be some of the big wins and
big hits in EDA for AI and machine learning.
Costello: It's better to be training wheels for machine learning than it
is to be machine learning.
Mo: Actually, data is the new oil, right, so Dean is doing it right
I think.
Anirudh: I don't think data and machine learning are like, exclusive, they
are connected.
Cooley: They just seem like buzz words to me, I'm sorry, it's just
it's...
Dean: Well, you know... machine learning... the original AI stuff you
know must have started 25... 30 years ago,
But the new Machine Learning and AI stuff is kind of crazy
because it's basically... it's neural nets, right?
And the crazy thing about ML is, people, it takes a lot of
training data.
They take a 100 million pictures of dogs, and they tell the ML
thing 'these are all dogs', and then they feed all the pictures
in and they train it and they don't really understand how ML
works but you know they kind of teach it and these neural nets
do this math and kind of come up with a result and then now the
next time they put in a picture it gives them a bit that says
'Oh it's a dog' or 'It's not a dog'.
And you know if you ask them "what do these 3 computations over
here do?" The guys says "I have no idea, you know the training
thing just taught it. That's what it was".
But, ML's got huge compute needs. Which is great for all of us,
because higher compute needs is great for the EDA industry and
the semi industry -- and it's getting a whole lot of people into
designing chips. I mean you've got Facebook, Amazon, Google,
and a 100 new start-ups designing neural network computation
chips.
I'm sure there's going to be some EDA start-ups that're going to
do some "neural net compiler" product that you know Anirudh and
Joe (Sawicki) and others will fight over buying or emulating.
But it's going to turn some of our industry a little bit...
I wouldn't say upside down, but it's going to have impact.
Anirudh: I remember I read this book, it was a long time ago - 20 years
ago or so, and I don't know if other people have read this,
talked about the Science of three P's, you know P.
So it had the first P was Science of Place. So that's geometry,
the underlying mathematics for it. And there's a lot of... if
you look at EDA, other places; a lot of geometry and geometry
processing, so that was the first P.
And the second P was the Science of Pace, so that's derivative,
that's calculus, that's Newton; 400 years ago. And these things
you know geometry is whatever, two thousand years ago right, and
the second thing was science of pace which is calculus which is
400 years ago, massive.
And then the third P was the Science of Pattern. And I think
finally there are some recent advances and we have used patterns
a lot, but there are advances that make pattern recognition much,
much more powerful. More powerful than humans can process.
So I think if you look at our industry or other industries, all
the three P's are important. I think if you have a new algorithm
for patterns, or science of patterns, first of all it can have a
fundamental impact that can last 400 years.
Secondly, there are definitely applications of that in all
algorithms we write; but it's not that we would replace PnR
purely based on science of pattern. You still have to do
science of place, science of of pace. So you will still use
geometry processing, you will still use calculus & differential
equations, and we will use machine learning. But the potential
is huge.
Sawicki: When you look at the whole thing of the hype, it gets ridiculous.
But what I can guarantee you is there are going to be significant
tools. Solido is already out there where you combine that...
and I love the way you (Anirudh) said it, it's the pattern
recognition. That ability to look at things as patterns,
classifications, driving behavior and then hybridizing that with
more traditional techniques; I guarantee you there are going to
be significant products out on the market this year.
Cooley: So why did Mentor buy Solido? Did they buy it for on-chip
variation or did you buy it for machine learning?
Sawicki: Yes. (laughter)
It's the on-chip variation characterization - it's enabled with
machine learning but it's not like... we have probably... I
know of at least like a dozen serious... probably likely to
result in product within the year or two -- projects with machine
learning within the company. That pattern recognition
capability, that ability to drive that kind of behavior, it is
bloody cool.
Dean: But you need data to train that pattern recognition. Big Data.
Anirudh: So you ask somebody, 'are you using machine learning?', it's like
asking them 'hey Joe are you using calculus?'.
Sawicki: Like C++?
Anirudh: Yeah, so it will be pervasive and there are a lot of applications
for it.
---- ---- ---- ---- ---- ---- ----
Related Articles:
Anirudh and Sawicki on why CDNS and MENT did the Cloud this year
Costello dissecting Montana, Rocketick, Palladium, Zebu, Veloce
28nm vs. 7nm, AMS, Virtuoso, CDNS Innovus vs. SNPS IC Compiler 2
Costello on EDA ossification, cloud, and RedHawk vs. GreenHawk
Join
Index
Next->Item
|
|
