( ESNUG 592 Item 08 ) ------------------------------------------------------ [06/27/23]
Subject: Tom Beckley on better/faster/cheaper chip design using Cadence AI/ML
The live DAC'22 Troublemakers Panel
Cooley: Tom. Aart has been doing really well in the air war for
artificial intelligence. I'm just seeing it everywhere. His
DSO.ai is everywhere.
Synopsys.ai is repeatedly mentioned in the civilian press like
the New York Times, Forbes, VentureBeat, making Synopsys look
like they are kicking ass in the AI world.
Does Cadence have marketing guys?
[ audience laughter ]
Beckley: Well, my first comment would be that Cadence Cerebrus with its AI
is competing very effectively and in fact, doing extremely well
in the marketplace.
And we would rather put our money into R&D to make our products.
[ audience laughter and clapping ]
I think all AI and ML (Machine Learning) is in its infancy.
I think 30 to 40 years from now, we'll look back and it'll be just
like printing press and the steam engine and electric motors.
It's going to permeate everything.
So right now, there's a lot of interest in AI/ML, but in many
ways, I think how it's being pragmatically used to really
change efficiency is different.
So, people are now putting their toe in the AI water. What we're
finding is that whether it's Cadence Cerebrus (digital PnR) or
whether it's in my space (full custom layout)...
... like Optimality, where Cadence has brought ML capabilities
into Clarity for EM (electromagnetic), Celsius for electrothermal,
and Sigrity -- what we're finding is that this design space is
too large. People can't comprehend this design space.
Once the designers pick their topology, we can help them find and
avoid local minima and maxima -- and optimize to that topology.
And by using the power of the cloud, and by using the power of
machine learning, we can give them a chance to do other topology
exploration, to do some "what if" analysis.
Cooley: Okay.
Beckley: I think that's where we're at, both on the digital chip side, and
on the kind of chip/package/board side of the equation. And I
think it's true for all of us.
Cooley: Dean. You've been pushing Big Data since 2014. Are you still on
that path or even jumped over to the AI/machine learning bandwagon,
just like everybody else?
Dean: Well, I'm a big fan of AI. And I've been investing heavily in AI
at one of my other companies, Eagle Eye Networks, in the video
surveillance base -- where I think AI is going to have a huge
impact analyzing video surveillance in a big way.
But the key to any AI is data. You got to have the data -- data is
training, right? AI is all about training, training, training. You
have to have huge data sets, and so the DDM and management of those
datasets becomes important. But the big data storage and database
to fill all the AI is really important.
So, we're still doing the big data stuff at IC Manage, to track and
understand what's being done with the design, what parts of the
design, etc.
Cooley: But Cadence Tom's tools automatically do it -- and probably Mentor
Ravi's does, too.
Dean: AI is going to be, in my mind, super vertically, super application
focused. You're going to have AI that can do this very specific
task.
The typical example is in surveillance video or in images or on
Google images, is we've trained this AI engine.
It's taken 14 engineers and 3,000,000 images. And we've trained
it to detect a dog or no dog. Right?
And so, if you need to know dog, or no dog, you've got this AI that
can tell you dog or no dog.
Cooley: Right.
Dean: And you know, and you might get your AI engine so that it can
actually recognize the super high, the super fancy stuff.
They've trained your AI engine so it can actually identify like
100 animals; or 100 different objects that's in an image.
And you think about that AI applied to EDA. Or to designing chips.
I mean, we face 1,000 different problems. And there's going
to be an AI thing that will be trained (or taught) that "here's
how you solve this particular problem."
What Tom was talking about is, this is AI that will give you some
guidance about how to partition your design. Okay.
Well, that's only useful after you've got a certain chip layout
topology started and you got some place to start.
But you know the AI on how to build an adder circuit -- the AI
will decide on which IP you should use, etc.... it's going to
be different AIs trained in a 1,000 different ways.
You need data for all of that -- and IC Manage is all about the
data.
Cooley: Ok. All right
Tony: And moving the data around.
Beckley: Yes, but I think we should be careful.... The chip data is the
customer's chip data, right? I mean, so that's how your AI is
going to build your foundation models.
Dean: Is it? Is the data going to be customer data? Are you
guys (at Cadence or Synopsys or Mentor) going to take chip data
from 100s of customers and create really smart AI that can
design a chip?
Beckley: Not unless we want to go to jail.
[ audience laughter ]
Dean: Well, so then, how are you going to ... So, are you telling me
that the customers need to do their own AI training, in order
to use AI in your tools?
Beckley: I'm saying that the customers own the large data sets.
Dean: Sure, of course... But how are you (at Cadence/Synopsys/Mentor)
going to train your AI?
Beckley: Of course, we have some of our own Cadence designs that we use,
and then we partner with the customer.
But to be fair, if you want to go across platforms and you want
to go across vertical segments -- that's owned data, right?
That's how they win or lose in the marketplace.
Dean: Yes. To get vertical.
Beckley: Yes.
---- ---- ---- ---- ---- ---- ----
Related Articles
Anirudh's 9 AI jabs at Synopsys and Ansys from his CDNlive'23 keynote
Reader snarkies on Anirudh 2023 vs. Synopsys.ai, ChatGPT, and chiplets
Dean, Anirudh, and Sawicki debate on Big Data versus Machine Learning
Join
Index
Next->Item
|
|
