SemiWiki.com - Podcast EP323: How to Address the Challenges of 3DIC Design with John Ferguson

Episode Date: December 19, 2025

Daniel is joined by John Ferguson, senior director of product management for the Calibre products in the 3DIC space at Siemens EDA. He manages the vision and product offerings in the Calibre domain for 3DIC design solutions. Dan explores the challenges of 3DIC and chiplet-based design with John, who describes the broad range of… Read More

Hello, my name is Daniel Nenny, founder of SemiWiki, the Open Forum for Semiconductor Professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is John Ferguson, Senior Director of Product Management for the Calibur products in the 3DIC space at Siemens EDA. He manages the vision and product offerings in the caliber domain for 3DIC design, so solutions. Welcome back to the podcast, John. Thanks, Dan. I appreciate you having me. Okay, let's dig in. As process geometries continue to shrink and die stacking becomes

mainstream, what multi-physics effects have emerged as the most pressing reliability concerns? Well, you know, there are a lot of different ways you can look at it. Ultimately, when we're talking about 3DIC space, it comes down to impacts of, power, thermal, and mechanical stresses. The challenge is that these things all have impacts on each other, right? You can think very simply about, you know, putting a blanket on at night, you get that pressure from the blanket that causes you to have extra heat. So these things are normal, and we tend to deal with them, but when it comes to chip behavior, particularly with thermal once you start adding that heat in there it affects the wires and how the wires

behave it also affects the transistors and the transistor behavior so if you're not careful and you're just you know taking chiplets you think you've got known good dye off the shelf i'll just put them in and i'll connect them and i'll be good probably that's not going to be the case, right? Because these things are getting hot. You don't necessarily know where and how that heat is moving and transporting across the greater system. And it's possible that the changes in the electrical behavior are going to knock you out of your electrical spec. Right. So if you're, if you're trying to do timing sign off, for example, and you haven't taken care of this, you haven't considered it and carefully figured out how you're going to deal with it.

it, then there's a good chance that you could actually miss your timing margins. So that's really the scare in the industry. The good news is, you know, all the EDA suppliers are working this, and there are solutions in the market that can help take care of it. That's a great answer, John. I do see that. There's a lot of problems with the chiplets and getting things integrated, especially with thermal.

So how has the integration of multi-physical? simulation evolved in response to the challenge of advanced packaging and the 3DIC technologies we just discussed. Yeah, I mean, now we're starting to see, okay, you need to have ways to check these things, right? And what does that look like? In the case of thermal, historically, there have been ways where we dealt with thermal. It was a little bit easier, right, because you're in a situation where you had maybe one die in a package. The chip designer would be done. They throw it over the wall to their packaging

team. The packaging team has some thermal experts, and they'll look at it and say, yeah, this is good. The problem there is that it's not as easy as that because your heat is really coming from the chiplets. That's where you have all the active energy happening, right? So you have to accurately model your chips in order to do this. This is an area, you know, caliber is very good at geometric processing. So that's not an issue. We know how everything is connected. We know how to pull all that together. But then you have to be able to take that and say, okay, now I see how this is connected. How do I from that figure out how or what the thermal behavior is going to be? To do that on our side, the way we do it is we bring in the Siemens flow therm solution.

We have built a tool we call caliber 3D thermal. It is a caliber offering. It is a single tool. But under the hood, we have a version of the traditional Siemens Flowtherm solver engine. So what we can do is we can say, okay, caliber goes in, takes a look at the chiplets or even the package geometries. It understands what the geometries are there, what their materials are, and it will then make a model, a thermal model of each chiplet within the context of the package, because you have to know what's surrounding it as well. And we'll say, okay, from this, I can now take that model. I can pass it under the hood into the flow therm engine.

Flow therm will come back and tell me, here's where your temperature changes are. We can do dynamic or static thermal analysis. We'll pass that, then back out, and we will present it to the user in a way that they're familiar with, right? Traditional caliber results kind of information, which ultimately makes it easier, particularly for a chip designer, right, that's used to, hey, I try to, I got to get my timing out the door, I got to get to final sign off, let's go. They know how to use these things. So it's really not putting an extra burden on them. That's the way that we're trying to look at things here. So, you know, then if I think about that, we can also say, well, there are also impacts of thermal on the,

Another problem, which is mechanical stress. As you heat things, they tend to expand, right? They get bigger. And so because of that, they're putting pressure on each other. Silicon doesn't expand much, but laminate does. And so the laminate will squeeze on the chip, and that'll cause it to kind of warp and potato chip a little bit.

And that ultimately causes issues with two things. One is it'll also affect, potentially affect timing because it changes the mobility in the transistors themselves. So you have to be careful about that. It also has the potential to cause you to have reliability issues, right?

Because this thing is warped a little bit, you know, do you really have confidence that your bumps and your pads and your TSVs, ESVs are all properly connected still. Sometimes they can stop, particularly trouble someone that happens in use, right? So it's a little bit of a scary situation. Similar situations on our side to solve those issues. Calibur 3D stresses our latest product in this domain. And it works very similar to our thermal approach, right? It is taking, the stress impacts that come presumably from your foundry or wherever you're getting your stuff manufactured. You take that information, we'll make a model of it. And then from that, we can determine what is the warpage,

what is the impacts on the electrical as well, which is something that a lot of the industry so far has neglected. So it's nice that there is a solution in that space at this point. No, that's good. That was good answer. So what role do accurate material property libraries and evolving modeling standards play in achieving predictive multi-physics simulations, you know, especially in two and a half D and three D I see structures? Yeah, it's really critical. I mean, we've had material properties in the 2D space, particularly for parasitics. You know, we know they're important. But now you've got to. to have materials that tell us how is this thing going to behave in the presence of heat?

How is this thing going to behave in the presence of warpage and stresses? So they're extra material properties. If you have the wrong properties, then obviously, you know, you're likely to have incorrect results in the end, and that's pretty risky. It's a little bit challenging here. You know, in the old days, we're building an SOC, we know how to put your cells together into a single piece of silicon. And you had fairly good confidence that when you're done, you're going to be okay. But if you don't have tools to do the thermal and the stress analysis, and if you don't have the accurate materials properties to properly capture those, that can be a little bit dangerous.

So that's also an area outside of caliber, but part of Siemens, we have abilities to do things like actually make measurements and essentially calibrate what the actual materials properties were. So that's another possible avenue of how to get that materials. The foundries don't like to share the actual materials that they use. They don't want users or customers to be reverse engineering their IP. But if you're actually measuring, it's really not telling you what the material is. It just says this is ultimately what the combined material behaves like. So it's a little bit safer in that respect. How are the engineering teams leveraging early multi-physics feedback,

you know, to influence architectural decisions and layout strategies or power deliveries? in network designs? That's very important. I mean, the big challenge or concern is if you go ahead and you design something without this and then you say, okay, now I'm going to go run sign off thermal analysis or sign off mechanical stress analysis, it's not sufficient, right? Because if you find something at that late state of the game, it's too late. You don't have time to go redesign the whole thing.

So you have to be thermally aware, power-aware, mechanical stress-aware throughout the design process. You know, there are different ways you can do that. There are iterative ways that you can automate multiple iterations. There are things like using AI as ways to do it. There's also the challenge that, you know, just like, in a 2D, right? When you're putting your chip together, you don't, everything doesn't come together all at the same time. So you have to be cognizant of, you know, I might have, in one situation,

I know I have a piece of silicon and I don't know what else is going to be on it, but I know roughly how big it is. And maybe I've got a good guess at the initial power, but I might just have one power for the whole piece of silicon. And it's not accurate, right? Right. But if you put that in context with some other chips that do have some stuff, you'll get some learning and it will help drive you in the right direction. Again, those right directions, that's also an area for AI where you can say, okay, you know, let's do that automatically. Let's have AI drive those iterations instead of human being and look for things like, hey, I've seen something that looked like this before. so maybe I don't have to recalculate the whole thing again. I can just reuse what I did before

and then I can do a more accurate assessment as we get closer to the final sign-off. So there's a lot to unpack there, but very, very important. You have some method of being able to bring in early information in the beginning and use that ultimately, whether it's manually or sort of, through automation to bring that to your final conclusion

that will give you much higher percent of success. Right. You know, it would probably help if you could share a real world example where multi-physics simulation uncovered an unsuspecting failure mechanism like thermal or electrical or mechanical that traditional verification might have missed. Yeah, we do see those. We've seen things like particularly thermal issues,

where you have dyes, big dyes, and maybe they didn't have quite the right materials properties, or maybe they had simplified or oversimplified some of that review, like they didn't have the detailed analysis of the chiplets, and they were only looking at it as silicon. You can be off on that. We've seen that in real customers. There were a few fairly big ones that had some, and yield hits as a result of that, you know, three, four years ago.

We've also had situations where mechanical stress did cause problems, right? Where it pushed things out of spec and the chip miss yield because of it. Usually the stress impact can be smaller, but particularly if you have sensitive analog IP that can really cause you some issues, but also even a smaller impact. sometimes it's just enough to push you out of the margins, right? So you have to be careful. You have to think of all of this as a holistic view. And what advances have you seen in tool interoperability,

you know, and data exchange that can help facilitate true cross-domain collaboration among design, packaging, and reliability teams? Yeah, it's tricky. One of the big ones for me has been the 3D blog. format out of TSM. That's a nice format. It brings away for all of the industry, all the tools in the industry, as well as all the manufacturing sites, to understand, you know, what is it that you're building here? What goes where? What is it you're expecting to happen? And so once we know what that looks like, then that allows all the tools to do whatever you want to do on it, right? You can run some of these multi-physical. tools you can do more simple traditional DRC or LVS or PERC style checks across them. It really opened up the door, I would say. There is a bit of a challenge.

I'll give a little warning that there's no Rosetta Stone here, right? So if you've got multiple people looking at the spec for 3D blocks, there's room for interpretation. and not everybody interprets it the same way. So sometimes we see, even though we have these capabilities that are built for interoperability, it's not always foolproof. It's not that bad. We're still relatively early in this stage,

but I think to the early days of creating GDS, even sometimes there were disputes between vendors and say, well, you know, GDS, it means this, no, it means that. In the end, we just talk it out, right? And we come to some agreement between them. It's not too difficult. But you may experience some issues in that going forward in the short term. Right. So, John, looking ahead, what industry development such as AI-driven simulation or digital twins or cloud-based workflows, do you expect to have the greatest impact on the future of multi-physics verification?

really all of those it's it's kind of interesting when i think about 3d i see i always think about a i because you know early on when we're talking about 2 and half d 3d what is it going to be we all thought well we're going to take the components that make up a chip and we're going to disaggregate them into tiny little chips and we're going to connect them up differently in ways that will give us benefits But, and there's some of that is happening, but the drivers today are very different. The drivers today are AI chips. These things are monsters, right?

These, if you were trying to make it on a chip, you couldn't because they're, they would have to exceed the size of a reticle. So you get multiple of these reticle chips that you're putting in there. They're big and they're heavy and they're hot. And they're really ultimately driving the way everything works together. So you need to have more compute power to run this stuff. You've got to rely heavily on AI. But it's AI that we're building, right? So it's kind of like a self-licking ice cream cone here.

You know, one hand feeds the other. Hey, thanks, John. It's always a pleasure to speak with you. And I will talk to you again next. Thank you, Dan. Always good to talk to you. Take care. That concludes our podcast. Thank you all for listening and have a great day.