( ESNUG 588 Item 2 ) ---------------------------------------------- [09/12/19]
Subject: Anirudh on each Machine Learning engineer is worth $10 million
DAC'19 Troublemakers Panel in Las Vegas, NV
Cooley: Anirudh. And you're adding "Smart Jasper AI with 2x faster
proofs. 5x faster!" And you know, smart proof technology.
(see CDNS PR 05/08/2019) Are you just doing more Press
Release stuff like everyone else? [laughter]
Anirudh: The good thing about Machine Learning is it's computational.
So, if you look at the software industry in the last 10 years,
most of the activity is in social media, which is not really
computational. Social media has great companies like
Facebook and Linkedin -- but machine learning is actually
computational.
So, when it is actually computational software, it's lot of
matrix multiplies and all kinds of numerical analysis -- which
is in the sweet spot of traditional EDA expertise.
I think what you will see is that ML will be widely adopted,
and the EDA companies are very well suited to do Machine
Learning in a fundamental way. That's why [at Cadence] we
are also doing a lot of machine learning, whether it's with
DARPA, internally, with Nvidia, or with Carnegie Mellon --
there's a lot of collaboration in machine learning -- because
inherently our [Cadence's] core strength matches very well
with machine learning.
To give you an example, I met this famous professor and he had
a machine learning startup. He's in this big famous university.
This is not in EDA nor design space, but it's a different space.
They were trying to assess the value of the company, because they
don't have revenue yet. The new metric for the value of a
machine learning company was $10,000,000 per machine learning
expert.
Cooley: What... ?
Anirudh: That's the new metric, right?
Cooley: Screw EDA, I'm going into machine learning!
Anirudh: So, I tried to compute the value of Cadence using that metric.
[laughter]
And you know, our market cap is about $18 billion to $20 billion.
It fluctuates. And we have about 4,500 R&D engineers. I think
all of them can do machine learning pretty well, so we are about
between $4-$5M per EDA machine learning expert.
So, we still have a 2X CDNS stock valuation to go.
We are working hard.
(laughter)
But what I want to tell you is ML is inherently computational
software. And if it's inherently computational software, it's
something that we in EDA can do well and there are a lot
of applications.
Cooley: How do you know you have an expert in machine learning, or if
someone has just put it on their resume?
Anirudh: No, what I'm saying is that, of course, we have to learn the
domain; but most of the things in machine learning are actually
not the algorithm itself.
You know, I don't believe that the algorithm is the really
complicated stuff. Most of the machine learning is modifying
your problem or mapping the algorithm to the existing machine
learning knowledge.
If you're doing verification, how do you modify it so that it can
make use of machine learning? That is the tricky part and the
applications are there in multiple EDA areas.
Naveed: I have a simple answer for that. We have an AI algorithm which
finds out if somebody is cheating on their resume.
Cooley: From machine learning or you just check it?
Naveed: Machine learning. We have machine learning way to figure out
who does machine learning.
Sawicki: I'm not even going to ask what your training set was. [laughter]
Cooley: Did you find a lot of cheaters?
Costello: Anirudh, I think it's true that the expertise of people in this
business could be applied to other places.
On the flip side, what are the best places here? Because as
you said, the core algorithmic basis for ML isn't necessarily
so complex, it'susually the data set that is most important.
Which areas do you think in EDA are going to be most fruitful
for applying a big data set to get much better results?
John, to your point, when people try to analyze and see how you
got to that point -- mostly you can't figure it out. Most
humans won't be able to be able to tell why it's a better answer.
Cooley: That's why Sawicki was saying, "Good luck. You can't figure it
out." Got it.
Anirudh: But there are certain areas which are very high return.
If you look at machine learning, of course, we have algorithms
which are already good for certain things such as matrix,
optimization, or Boolean analysis.
But there is verification, for example -- logic and functional
verification is a very difficult problem. It requires a lot
of human intuition and knowledge. It's NP-hard or NP-complete;
you know, all the complications with verification.
I think verification, first of all, is a beautiful area for
machine learning and we already did some work with the Jasper
on it.
But right now, you run verification for 6 months and still
don't know whether you are finished with verification or not.
So, overall verification closure is a great area for machine
learning.
Of course, layout-based patterns is another good area. Whether
that's applied to DRC/LVC or it's applied to place and route.
So, I think these are definitely two areas in EDA which are
ripe for machine learning.
---- ---- ---- ---- ---- ---- ----
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
Join
Index
Next->Item
|
|
