Cadence Design Systems, Inc. (CDNS) Earnings Call Transcript & Summary

We'll get started. Now good morning, everyone. I'm Stacy Rasgon. I'm Bernstein's senior research analyst covering the U.S. semiconductor and semi-cap space. And it's truly an honor to have our guests here today. Dr. Anirudh Devgan, the President and CEO of Cadence. Before I start, I want to mention if you have questions you'd like to ask during the presentation. You should have a QR code in your program that links to our pigeon hole form where you can submit those questions. We'll have time for Q&A at the end. Cadence has also asked me to read their safe harbor statement, so I will read that now. Today's discussion may contain forward-looking statements, including Cadence's outlook on future business and operating results. Due to risks and uncertainties, actual results may differ materially from those projected or implied in today's discussion. All forward-looking statements during this meeting are based on estimates and information available as of today, and Cadence disclaims any obligation to update them. Great. With that, let's dive in. Anirudh, I'm so happy you're able to join us -- thank you so much for coming. I want to start at very high level. So this is primarily a generalist conference and not everybody here maybe is deep in the weeds on semis. And some of them may not know exactly what EDA is -- can you just talk a little bit just what is cadence? What is ED -- what does it stand for? What does it do? Where does it fit into the semiconductor supply chain? And where does Cadence play within all of that?

Yes. Thank you for the question, and thank you for your interest. And my kids don't know what EDA is. Actually, they barely knew that there are chips in these things they were using. But at least now everybody knows their semiconductors and phones and TVs and cars. But what we do, what Cadence do at the highest level is, and we basically make products to design semiconductor chips and electronic systems. So it's mostly software products. We have some kind of supercomputers to accelerate software development. But this is to design and then all the chips that you see around you, and we are privileged to work with all the leading companies in the world in all kind of verticals, in all geographies. So that's it in a nutshell.

Yes. So just -- I guess it's sort of a software environment that's used to design the chips and super computers used to [indiscernible]. That's correct, so simulating those designs before you actually put them in silicon and realize something is wrong.

Exactly... So the thing with semiconductor is that it has matured over the last 40 years and of course, going to a resurgence now, okay? But they are -- I mean, at the simplest level, they are very, very complicated, right? So there cannot be designed with like -- it's not like your architect can draw like a 5-bedroom floor plan of your house.

They used to do it by hand back in the days.

But now there are like some of these chips will have 100 billion, 200 billion transistors operating at 3-nanometer or 5-nanometer. So all of this is done by mathematical software or engineering, what we call computational software. So we will have a high-level description, then we do synthesis optimization simulation. So it's a pretty involved area. And so all our customers will work our software to design, whether it's CPU, GPU or phones or cars chips.

Got it. Got it. And who are your customers? Like is it just every semiconductor vendor on the planet buys your tool sets or?

What we like to say is that -- and also Cadence is the most diversified EDA or design software company. So we have software to design digital things like GPUs and CPUs, software to design memory and all the big memory makers, software to design analog. So what we like to say is that almost any chip design the world today use some form of Cadence software. So we are lucky to work with all the leading semi companies and some of them are public, of course. And then also all the leading system companies that are designing chips because the other big thing that happened over the last few years is these big system OEMs, the phone makers or the car companies, and this is well known, right? Like Apple has a whole line of MCDs and chips, of course, Tesla is designing their own chips, all the data center companies are building their chips. So what happened in the last few years is not just working with the semi companies, which are great, and they themselves are becoming system companies like NVIDIA is truly a system company. Qualcomm, truly a system company, Broadcom. And then the system companies are also doing silicon. So that adds to our customers.

Got it. How does working with those system companies differ from working with? Like do you have to do a little more handholding upfront? Or like I'm probably grossly simplifying, but.

No, these are remarkable companies. I mean -- and there's a range of capabilities. I mean some of them are best-in-class, right? So I mean they don't require any handholding at all, right? But some of them, we do have some value-add services. It's not our main business, but sometimes we do provide some value-add services, especially to system companies. But yes, I mean the -- but now they have -- a lot of them have really remarkable teams. And also they have an advantage of -- see, when you're a system company, the advantage is you know this anyway. They have a software stack, otherwise, we wouldn't be. So there is a lot of optimization that can happen with the software stack and the silicon. And so things like emulation become more important in system companies. So what emulation does, which is our hardware business, we are mostly software. But what happens is when you are designing at the highest level, so for people -- I mean, for those who are familiar, this is like basic stuff. If you're not familiar, you just -- so what used to happen in the old days and pardon my use of hands is that you would do like silicon development, right, will finish, it would take like 4 or 5 years in the old days. And then you would do software development, that would take like 1 or 2 years, and then you will release the system, which is hardware plus software. Now this is too long now. So they get overlapped. So -- and this perfect example of all the latest chips there. You do the hardware development and software development in parallel. So when the chip is finished, it just comes out within a few months, and it has all the software running on it. So the only reason that's possible, and this reduces the cycle time because if you want new chips every year, the chips take time to design and software, you have to overlap it, okay? So the only reason this is possible is that when you're doing software development, you don't have a chip, but you can emulate it. So we sell these supercomputers...

They can run the software on the emulator?

Yes, exactly. So all these system companies and these big companies like NVIDIA is a great -- is a development partner of palladium. So they will run.

Let's talk about NVIDIA.

Because they have talked about it publicly at Jensen and NVIDIA. So which we really appreciate. But so NVIDIA will run software stack, whether it's windows or CUDA or Android or the same thing with the system companies well before they have a chip. So -- and these emulators will run that software. I mean, slower than real life, but maybe thousands times faster than if you would run it on a general-purpose CPU. So they become essential to the design of semiconductors and software. So the other thing that is different with the system companies. I mean, semi companies, a lot of them have software anyway. But the system companies definitely have software.

Got it. Got it. How do they used to do the [indiscernible] FPGAs or something? Or what kind of they [indiscernible]?

I think, yes. It was just more cumbersome -- you can do it with running it on -- in the old days, you would do it by running it on just regular x86 CPUs, but they are too slow to really do a software bring up and verify. So the last -- we have done this for like now like 20 years, but the importance of these emulation systems became more and more critical as the chips get bigger...

Do you think that's like one of the bigger development issues that the industry has faced. I mean, because there's a bunch of stuff. You have an increasing complexity of process technology like multiple patterning, shorter design cycles that you mentioned. Chip development in general is getting more challenging and more expensive. And so what other sort of like, I guess, specific innovations as Cadence,have they brought and are they bringing to the table to help reduce those cycle times to reduce that design. Because I've seen some of these charts, which shows like design costs. I mean I don't know that I believe the charts because I think if they were real like we wouldn't be designing any chips at all. There's no way.

Well, people like to project these just like right now, the power demand will go through. So I mean the design costs have gone up to some extent. But I think what is important to remember, which sometimes is missed in this, okay, is that those have to be normalized to the volume of silicon. Because this is upfront design cost. So if you're selling like 1,000 chips or systems or $1 million or $10 million, that's a very different economics. So the good thing is that because of silicon proliferation into everything, the volume of things are going up a lot. So the costs normalized by volume actually going down. That's the reason that all the system companies want to do things is because they have high volume also, whether it's data center AI or phones or cars, right. Now -- but we are always trying to make things more and more efficient, and it's the history of EDA. In the last 20 years, we have improved design productivity by 100 [indiscernible] because remember, in the early 2000, it would take like a few hundred -- 500 people to design a chip in like 5 years. Now you can do that in 6 to 12 months with probably 40, 50 people. So -- it's still more to be done, but the history of EDA, the A stands for automation, right. And of course, the big automation now going forward is AI.

We'll get there [indiscernible].

So that's another level of automation.

Got it. I guess just one more question just on the broader market. How do we think about EDA growth? Is it just does it grow with long-term semiconductor? Or is it like semiconductor R&D a percentage of that? Because I always sort of looked at it as kind of the latter. It was almost like -- and it's weird because it gets into the pricing discussion, too, right? Because I feel like historically, you were not really negotiating pricing with -- it was mostly with the procurement folks, right? It was like we spend x on R&D, and we spend y percent of that on EDA and like that -- is that how it is? Is that dynamic changing? Like how do we think about the drivers of the industry driving growth of the EDA industry?

Well, one thing I would like to say is that just like the value of semiconductors have gone up, the value of EDA has gone up, especially the more advancing what it is. And you can see it in system companies too. I mean, the value of silicon and system companies has gone up. And then in our EDA interaction, we are having more and more interaction at the CEO level rather than what you mentioned in the past. And because it becomes so essential to getting more kind of realization of value -- because they're already investing billions of dollars in these node transitions from 5 to 3 to 2. So you had to make sure EDA, the design part is also as -- so that has gone up for sure. But we are, we are part of semiconductor R&D spend. Now the other good trend in the last like 10 years is EDA as a percentage of R&D has increased because there is more and more need for automation. And then we have new customers like system companies. So there also -- I would say the 3 big things would be I think the importance of EDA is higher in our customers and also in the foundries in all the major foundries. And then the -- we are a higher portion of semiconductor R&D. And thirdly, I do think that this new system company is adding more R&D. And even semiconductor R&D, even though semiconductor had like a difficult '23, and I think hopefully '24 and '25 is better. There is realization because of all this competition that they do need to invest in design in EDA.

The system players value more what you do than the semiconductor is because they're newer to it?

No, we love both our customer sets. And I think they're all very capable there. And there is also movement between them, right? Some of the people so No, they are both fabulous and we have great -- they have the opportunity to work with [indiscernible].

Let's talk about AI. Let's -- So I mean, you certainly talked about improving efficiency, design efficiency. And where does AI drive that? If you went from 500 engineers in 5 years to 50 engineers in 2 years? Like does it go to 5 engineers in 1 year? Like where are we driving efficiencies with this? And maybe even sort of like a broader overlie what does AI and EDA like mean? Like how does that what does it actually look like?

Exactly. Yes. So give me a few minutes. This is a big topic. Because the one issue with AI is that everybody calls everything AI so you don't know what is real AI and what is -- so okay, let me back up a few like give me a couple of minutes to what I think what AI is. Because one Webster dictionary definition of AI is like any automation that humans could do is be AI for that metric, we are doing AI from when I was in kindergarten okay? So that's not my definition of, okay. So now to me, AI will have 3 kind of phases of deployment. So if you let me say that first, then I'll talk about. So that's true for any technology and then people say about AI is like the Internet, okay. I think it's more fundamental than the Internet, but we can take Internet as an example, okay. It's more like if you really want to get into it, it's more like calculus or electricity, okay. So -- but the first phase of AI is always -- our new technology, always the infrastructure, which is what we are in now and it's probably next several years. And silicon is a key part of that infrastructure. So Cadence will participate in that. All the AI companies making chips and systems. So this is a thing that's really different from 6, 7 years ago where semiconductors was like you said, maybe maturing industries.

Yes, I was joking, we did a piece in 2015, a big one, it was called playbook for a maturing industries. in 2017 or so.

So I think what is this whole infrastructure build-out, which we are far away from and that they use our products, regular products, not just AI products, okay? The second phase of a new technology, second phase of AI is applying AI to your products, whether it's search or Far Point or EDA. And we are, of course, doing that for at least last 5, 6 years, including Gen AI. And then the third phase of AI will be like new applications that will emerge that never existed before, just like what happened with Internet, with Facebook or social media. So I think AI will go through these 3 phases, first of all. So -- and we are participating in all 3. And the first one is with the infrastructure build out with our customers. whether they're a system or semi. The second one is what you asked about, like, can we apply AI to ourselves. And of course, we can in a big way. And so -- and the reason that what AI can really do for EDA, which EDA never did the really new part of AI. So which is -- so if you look at EDA, we write all this complicated software, when you run it, it takes like 1 or 2 days to run some of the most -- this is like design software. So you give it some RTL description, then it will do the synthesis, place and route, timing, performance, tell you how fast it runs okay. And these chips will have like hundreds of blocks. So each designer is running, let's say, one block.

What's a block, sorry.

Block like a number of transitions. So it could be like a camera unit or a CPU unit. So typically, it's like few million gates or like tens of millions of transistors. So one designer is designing like 20 million, 30 million transistors using our tools, okay. But what EDA -- and these are like complicated things that run for a few days and give you a beautiful answer. But the design process doesn't happen in 1 or 2 days. We will be testing with any design process that's iterative, so what happens is you run these tools, then you change something, you run it again, you change something, you run it again, and that's human intuition. So the driving of the tool is done by the EDA.

Is that really how they optimize it? Is it really just the intuition of the engineering or the.

Because the tool is doing a lot of work, but like some of the of course, is intuition.

I'm a dumb process engineer, so like the EDA goes over my head.

The intuition of the engineers are critical. Also the architecture, this is what we want to do. This GPU should have this much this CPU should have these features. We want to go to TSMC 3-nanometer, we want to have this power level. So there's a lot of like high-level design, of course, that is the bread and butter of our customers. But then to implement that, you -- it's a iterative process just like in anything, right? You -- so what EDA never did was that we were very good at what I call single run, right? You give it a input, it will give you a beautiful output. But when you go to next day or 2 days later and restart that, it has no knowledge of what happened in the previous -- and this is true, in general, in a lot of things. Like when you open a new PowerPoint, doesn't know what you like in the -- so now it's not that we didn't want to do that. We always wanted to do that. but it was mathematically not possible to do it because you are transferring knowledge from one rent to another. There are some ways to do it, but they were too expensive. Like so what you're doing is you're searching the design space. I'll give you an example, like if you're doing CPU design, you're searching this design space and you want to get a best power.

How many like dimensions or variables are you like searching?

It could be -- I mean, like even at the -- now when you design the things, it's like a huge number, like I said $20 million, $30 million, sometimes billion. But even at the user level, okay, this is at the tool level is doing millions of variables and billions of variables. But even at the user level, it could be something like 15 to 20 variables, which is a lot, okay? In one case, so that's difficult for any user to do. So that's why the human intuition is important. Okay.

How many iterations are typically required...

A lot of iterations -- so what happens is you said, "well, why don't we do it mathematically. The classical way to do it, now some people may call it AI, but the classical now statistics is called , right? But there is a classical way in statistics -- by the way, this is not true for companies. It's also true for universities. All my CMU now they have school of AI and not school of computer science. That's my alma matter MIT is all -- everything has to be AI, what used to be computer science and that is now AI. So in the classical statistics, the way to do that is design of expert sorry, taking a little longer to adjust to it. So that -- the problem with that is that it would take millions of runs, even for '17 variable. So that's infeasible. So then the way was mathematical way was infeasible. And then the human way, of course, works, but it took like 6 to 12 months.

You may not find like a global optimal level.

But still it was great. I mean, it still much easier than 20 years ago. And then there are some other mathematical ways which they don't work. But AI, there's reinforcement learning and AI. The beautiful thing is that it can build a model of anything. So you can actually do that using AI. So for the first time, this is now real AI. We can actually do this knowledge transfer from one run to the next run and do this workflow automation. So instead of the user just running one time, the machine runs. And so in this case of the CPU example I was giving with 50 to 20 variables. Now it didn't take 4 million runs, but it takes like 200 runs or 100 to 200 runs, and it can do the design, okay? And then we can paralyze those runs because some of these things are parallel. So what -- the big thing is what used to take 1 or 2 days in one run now can be done in 1 or 2 weeks instead of doing it in 6 months, you can do it in 1 or 2 weeks with AI, with AI with 200 runs. So there is a lot of benefits to that so one benefit is that, of course the still [indiscernible], they're not done in 1 AI run but still could be 5x, 10x efficient, more efficient so far. And that's a lot, okay? So what that means is that 1 engineer instead of doing 1 block can do 3 to 5 blocks so that's the real productivity benefit of AI. And the second benefit, which is -- can be even more impactful is that, that optimal is better than what a human can do. Because in a lot of cases, we are able to design chips, which are lower power or faster, and there are a lot of examples, sometimes 8%, sometimes 10%, sometimes 15%, which is huge because you're going from 1 node to another node, the improvement is normally 5% to 15%, and you are getting that by better search, better optimization. So those are the 2. Now what is the impact to our customers and our business? So now the typical thing with AI automation, let's say this is another way of doing automation. And it's very -- now we are all top 20 customers, all of them are engaged with our AI tools because this is a real benefit. This is not like marketing AI to me, okay? This is real AI. That's why I took a little time to explain what it is. So the benefit is...

Genuine use case, right?

Yes, this is the other thing that AI, so you need some level of automation already to apply AI on top. See, if there is no automation in the environment and you just apply AI models, it's more difficult to automate. Whereas EDA has done decades of automation. So we are actually a pretty good use case. So then what happens is then the other thing with AI is, okay, you have so much automation that you will need less people or less tools, okay? So first of all, that's not the history of design. And the reason for that is that it -- there's several reasons, but 1 big reason is that's assuming that the workload is constant and then you apply automation, then you need less work or less people, okay? That may be true in some things. I don't know how many lawyers you need to process a document or whatever, okay? And not to pick on lawyers. But if you look at chip design, the workload is going up like this. So the biggest chips right now are like 100 million -- 100 billion transistors, and they're widely predicted to be 1 trillion by 2030, okay? So that's 10x bigger in size than the verification complexity, software complexity. So the design complexity from '24 to '30 will go up by at least 30x, 40x.

Is that scale linearly, I feel like it should scale more than linearly if that...

No, I'm coming to that, I'm coming to that. So let's say, the design complexity goes up by 30x, 40x, okay? There is no way our customers or their system a [indiscernible] you want to hire 30x, 40x more designers or engineers. They're not even graduating, 30x, 40x. But I think they will still hire more because they will be more silicon. But I would -- if I were to guess, I would say they will hire 2x, 3x more. That's reasonable in the next 5, 7 years. So then still, there's at least a 10x productivity gap that has to be filled and which we have done in the past also. So that's what now AI can provide this gap. Because even if something gets really nonlinearly faster or better, there is still some design iteration process. So actually, to really affect like 5x or 10x productivity in the real measurable design process, requires component level to be even more efficient. So I think 5 to 10x is a very good goal. Some people may claim higher, but I'm talking about real end kind of output. And that can -- and that's a very -- the other good thing about semiconductors and application of AI is because the workload is going up. So everybody wins. The customer wins because now they have to use more automation and more compute, but they can get better designs and not invest as much in headcount. And we win as a provider and then, of course, the customer wins because they have better solutions here.

Got it. No, that makes a lot of sense. It does sound to me that if 1 believes that, I mean, the EDA should be taking a bigger percentage. I mean it has, it sounds like it should take materially more. Does this help with -- I've always wondered about pricing in this industry, right? And it goes back to my comment on well, we spent X percent like that said. Does this give you -- I mean this sounds like a lot of value-add, like how much of that can Cadence capture, like as you're delivering this? How do you think about that?

Well, the way I look at it is we just have to provide value to our customers. Everybody producing a product things they're not valued and everybody buying the product thinks [indiscernible]. So we want to make sure we provide enough value and we save our customers enough things and we capture part of that. So I think I just want to make sure we -- we provide value to customers, which we will. And then they will keep some of the savings, and we should get some of that.

But everybody wins to your point...

Exactly. Yes, because we cannot -- we are not -- that's not our culture to just go on the pricing. We want to make sure, and we have a lot of amazing customers helping us in a lot of ways. We have to deliver value to them. And they're always fair to us. We get our part of the share of the gains, yes.

So let's talk about like what you're doing with NVIDIA. So NVIDIA has -- I mean they have a whole library, et cetera. I think it's called. They have their new Palladium emulating in I'll let you tell the story, but Jensen has made some interesting comments on that. What exactly are you doing with NVIDIA? How are you guys helping each other around these lines?

Yes. We have a great partnership with NVIDIA and NVIDIA, of course, is a remarkable company. And even more so in the last few years, our partnership has increased significantly. But we have a history of working with them more than 20 years. Now one of the -- and I think as Jensen has said that they use almost all of our products, which we appreciate, whether it is packaging tools or digital design tools, analog tools, Palladium. Now 1 -- and sometimes, we codevelop products with them. So for example, Palladium, which is this emulator, NVIDIA is 1 of our main development partners. They will kind of work with us years in advance based on what they need, the size of the chips and the requirements.