3D IC Front-End Design 3D IC Podcast Video

Show video

Welcome to the Siemens EDA podcast series on 3D IC  chiplet ecosystems brought to you by the Siemens   Thought Leadership team. In our recent podcast on  3D IC topics, we talked about chiplet ecosystems,   the design workflow needed for 3D IC, the  current state of 2.5 and 3D SIP design flows,   and the evolution required to support the  design community to develop these 3D IC-based   devices. Today, we will discuss some of the  front-end architecture, RTL level design,   and verification aspects of the 3D IC flow. I’m  pleased to introduce two special guests today:   Tony Mastroianni, who is the Director of Advanced  Packaging Solutions at Siemens Digital EDA,   and Gordon Allan, Product Manager for Verification  IP solutions, also at Siemens EDA. Welcome back,  

Tony, and hello, Gordon. Thank you both for  taking the time to talk with me today about   3D IC front-end architectural aspects. And before  we dive into the discussion, would you both mind   giving our listeners a brief description  of your current roles and background?   Sure, John. My name is Tony Mastroianni, and  I’m responsible for developing our 2.5 and 3D   IC strategies and workflows at Siemens EDA. My  background prior to Siemens has been primarily   in IC design, and mostly project management over  the last several years. I was involved in advanced  

packaging flows at my previous employer, which  was a fabulous semiconductor. And there we were   developing very complex integrated circuits  for various customers. I was there about 18   years. In the last three years or so, we  started getting involved in 2.5D designs,   incorporating HBM interposers, as well as a  design where we actually split a chip into two   dies that were integrated in a 2.5D package.  While I was there, it became apparent that   our traditional design methodologies needed some  major enhancements. So, for the last three years,   I was working there, developing our new integrated  packaging and IC design flows. Started about a  

year and a half ago at Siemens, as I mentioned,  working on the 2.5 and 3D workflows.   Hey, John, thank you for the introduction and  for inviting me to talk with you today. My name   is Gordon Allan, and I’m the Product Manager for  our Verification IP portfolio here at Siemens EDA.   My background is in SOC design and verification,  stretching back to the early 1990s, where I gained   the broad knowledge from spec through to silicon.  But my main area of expertise is verification.   I was one of the architects and authors of  accelerating UVM. And I’ve been with Siemens   EDA for the last decade, bringing solutions based  around SystemVerilog and UVM to our customers.  

Thanks to both of you for sharing that stuff  with us. Let’s get into the front-end design   topic for 3D IC. We’re hearing a lot of talk  in the industry about these technologies   around 3D IC and the technical challenges.  And many think these technologies and flows   are dominated by physical aspects, packaging  technologies, thermal stress, mechanical stress,   all of those great problems that we have to solve  and they were putting flows into place for. But   what about the front end of the IC design  process, chip architecture, and RTL design,   and RTL verification? Is that even relevant  to this? And why are we talking about this?   Yes, John, this is very relevant. Traditional IC  design scaling has been accomplished primarily   through IC technology scaling over the past  30 years or so. This process is referred to  

as Design Technology Co-Optimization. But as the  IC technology scaling has dramatically diminished   with Moore’s law, a new process named System  Technology Co-Optimization is extending this   design scaling. This System Technology  Co-Optimization, referred to as STCO,   is about enabling architectural and  technology trade-offs early in the   system design process to achieve high-performance  cost-effective solutions and reduce timeframe.   Predictive modeling is a fundamental component  of STCO that leverages high-level modeling tools   during the planning phase to home in an optimal  solution. 3D IC design implementation requires   co-design and co-optimization workflows. But  before you jump into the back-end packaging flow,  

this decomposition needs to happen at the  architectural level. But how do you do that?   How do you even know what different options  are available? And whether they’re valid for   your requirements? And if so, how do you home  in on the right microarchitecture, where you’re   starting to partition things into blocks?  Now, with 3D IC and chiplet ecosystem, those   blocks may be blocks within an ASIC or they may be  separate chiplet blocks integrated into a package.   These templates could be off-the-shelf devices,  or they could be a full customer basic design.  

So, to determine which microarchitecture is  best for your application, you need to do some   high-level predictive analysis quickly and find  an architecture that meets your specific product   requirements. Those requirements or priorities  may include power, performance, physical size,   or footprint of the product you’re building,  nonrecurring engineering cost and the unit   cost of the devices you’re building, and time  to market. So, some customers will be designing   at the bleeding edge of technology, while others  will be looking for more of a “Lego block” type   of approach where they can just put these chiplets  together and snap them together like a Lego block   to optimize the overall cost of these  complex system and packages. Now,   going the chiplet route needs to be considered  as an additional design paradigm.   So, let’s look at those multiple things,  actually all of the usual concerns for   any normal single-die SOC project, architectural  analysis; and then factoring in manufacturability,   that would be test, package, thermal, stress  analysis; and the functional verification of   your design your interfaces, memories, processors,  integration; and then there’s physical assembly   and verification, which is floor planning, timing,  bandwidth, signal integrity, and power. And we’re   looking now at async interposers, packaging, and  PCB techniques. So, for the microarchitecture  

analysis and optimization, the idea is identifying  your die-to-die interfaces and your typical   components and the associated data interfaces,  capturing those viable design scenarios,   and then run your functional simulations  and do some initial high-level analysis of   the power of thermal throughput. So, this is the  predictive analysis we need to do upfront. There   is an analogy here, I think at the 3D package  level, it’s like doing upfront floor planning.   And just as we must do SOC pad ring design  today, and now we’re talking about synchronizing   multiple pad rings within the package. And in  my world, which is RTL functional verification,  

there’s the need to have functional verification  of the off-chip, and die-to-die interfaces with   bandwidth and latency, overall design behavior,  performance, throughput, and so on, all of these   need to be measured. Ideally, we’d like to be  able to do as much of this analysis as possible   at this initial high level before the serious RTL  development and integration begins. So, the idea   is we will go in and start selecting the packaging  technology and then the chiplet mapping and   interconnect, make some choices there, understand  that high-level floorplan and pad rings. And then   we would do the mapping to target implementation  where we can look in more detail, verifying our   PPA. And the tool flow can enable more detailed  floor planning and some rough routing of some   of the channels to do some preliminary  signal integrity analysis and so on.  

And there we’re getting towards the collaboration  between the system designer or RTL architects,   and package architects and design teams who  would assess those deeper back-end concerns.   Okay, you’ve got me convinced it’s necessary to  get started with the 3D IC process upfront. So,   let’s look at the front-end design and  verification aspect of 3D IC. What do  

the SOC architects, design leads, verification  leads need to know as they take steps into the   new flows for their upcoming projects? There are several areas here that we can   discuss. One is architectural decisions  around the packaging partitioning reuse   that affect the ICs functional architecture.  The second would be interface connections;   how to communicate from die-to-die and how to  design that communication channel. Third would  

be interface verification; how to integrate and  verify all of those die-to-die connections using   standard protocols and memory interfaces. That’s right. In any 3D IC project, certain   packaging and partitioning decisions will be  made upfront from fixed criteria and experience,   and others will be decided during the course of  the architectural exploration and definition phase   by evaluating several options and choosing  one that meets the evolving requirements.   Still, others will be deferred until the  project has sufficient technical unknowns   resolved to finalize these decisions. You could  see that the introduction of this new technology   has an impact here. Actually, it brings a great  opportunity that was not previously available   to chip architects. But with that, comes a need  to consider the impact on the whole design flow,   manufacturing flow, and associated cost. Right. Packaging and partitioning is hard  

enough in a single-die SOC flow. But now we’re  introducing chiplets as solutions for reuse,   for integration, for handling disparate power  or other technology aspects. The chip architect   needs to evaluate and narrow down the options  early on in the process. The key messages here,   you’ve got a lot more tools in your toolkit and  one more degree of architectural freedom that you   didn’t have before. It’s a problem but there’s  room for optimism here, it’s a good problem to   have. We believe that 3D IC is a topic that every  front-end design lead or SOC architect needs to be  

informed about today. In fact, there are two  new things just announced in this last week,   which are relevant here and which will help  change the game for 3D IC technologies. I   want to mention them both during our discussion  today. First up is the new industry consortium   announced on March 2nd, the Universal Chiplet  Interconnect Express standard, or UCIE for short,   rhymes with PCIe. This has several major market  players bringing together the best of PCI Express  

Gen 6 technology, and Compute Express Link or CXL  technology into a proposed interface standard for   die-to-die communication. UCI Express will  leverage the PCI Express Gen 6 interconnect   and physical layer to provide a known quantity,  robust, fast die-to-die interconnect. Letting you   as architects make partitioning decisions your  own way, knowing that the interconnect can be a   solved problem. We anticipate that this will  help with the expected ecosystem of chiplets   available for integration and packages. And of  course, UCIE joins a list of other contenders for   that die-to-die interconnect, XSR, USR, AIB, and  others. But we really see its potential here.   So, clearly, there are a lot of concerns to juggle  early in the chip definition and development   project before we go off and write RTL and  verify it. You mentioned that in some respects,  

this is just like a normal SOC design process,  but just more. Can you comment on some aspects   that are particularly interesting or enabled by  the adoption of 3D IC and chiplet technology?   One area of interest, I think, is functional  safety and redundancy. We see in certain markets,   like automotive, redundancy is a big deal because  it’s a harsh environment. So, you need chiplet   redundancy in hardware and software systems  if that chip is going to drive the car. So,   they’re going to use different architectures,  so if one fails, they have a backup. But even   on the interconnect and 3D IC, these things  are like tiny little bumps on a substrate or   interposer. There are a lot of vibrations going  on in the car, so there’s going to be some finite  

risk of mechanical failures and so on, on those  connections. This necessitates smart technologies   baked in with Redundancy and Repair techniques,  or R&R for short, that are supported today,   for example, with HBM memories and in some of  the other interconnect protocols that’s going   to be required for automotive applications. That’s a good point. In harsh environments,   such as an automobile, the internal die-to-die  interconnect on the interposer, which connects   the triplets together, can fail from either  electrode migration or chemical stress. And  

for these types of designs, redundant routing  channels can be deployed. So, test hardware and   methods are available to detect and actually  repair these defects by rerouting the failed   channels to redundant channels that are designed  into the interposer. This approach can also detect   and repair memory defects. Additionally, as  you mentioned, redundant internal blocks or  

even chiplets can be deployed in the system, and  then swapped in on the fly during the operation   of the device if one of those components  or devices is not functioning properly.   It’s interesting what multiple chiplets bring to  this problem. Some aspects of functional safety   should be considered upfront at the architect  stage. There are multiple levels, the concept   of redundancy within the package or within the  die, as well as with the package multiple levels   of potential redundancy. So, a chiplet approach  could help deal with your multiple redundant   elements which are now often separate chiplets,  so they’re unaffected by point failures in common   mode concerns like power, thermal, mechanical.  If one of the chiplets fails, the other two   are going to survive and still drive the car. That’s a very interesting conversation, guys. So,  

what are you, in EDA, doing to help front-end  chip design and verification teams who are   looking at partitioning, and interconnect, or  architectural definition, redundancy? That is,   what flows and solutions do you have here? We’ve heard a lot in other podcasts about   different parts of the floor. My own specialty  is in verification IP. We provide solutions for   PCI Express Gen 6 for Compute Express Link,  CXL, for advanced memories, DDR5 and HBM3   memory interfaces. And we have customers  across multiple markets, processor makers,   memory leaders, SOC makers, aerospace and  defense leaders – they’re all interested in   this technology. One of our areas of strength is  the automation we provide to auto-generate test  

benches and let design verification teams get up  and running in minutes. This kind of productivity   enables architectural exploration of the sort  that we’re talking about here, and will give early   confidence in the solution space that architects  are looking at. And you can be sure that we will   be providing solutions for the emerging 3D IC  interfaces such as UCI Express as part of Siemens   EDA overall end-to-end flow for 3D IC. As an industry, we’re making a bet on this,  

in that we can tip the balance financially  in technology and help customers achieve more   complex designs by disaggregation, decomposition,  by deploying STCO. And with this disaggregation,   we can have multiple teams, each working on their  respective domains, making verification easier to   divide and conquer. So, we help those teams make  their architectural trade-offs and their early   experiments and exploration. As Gordon said, we  can provide verification IP and workflows that  

help to automate the rapid generation of these  scenarios for exploration. And then, ultimately,   the rapid generation of more detailed scenarios  once we’re honing in on a chosen architecture, so   that we can go into the packaging implementation  flows with all of this upfront design completed,   including the architectural design and the  upfront RTL verification, knowing that we   have a solid solution that’s going to meet our  functional goals going into the packaging flow.   So, earlier, you mentioned that there  are two pieces of major news this past   week. We talked about the UCI  Express already. I’m curious,   was the other one related to  Apple’s new M1 Ultra processor chip   announcement? Is that causing the stir? Yes, indeed, it was like a one more thing,   a major 3D IC announcement from Apple that we’ll  talk about. But first, the audience listening to   this podcast might be in all kinds of different  end-use markets for their ICs. And the question  

is not which markets are relevant for 3D IC, it’s  becoming more like which markets are not relevant   for 3D IC. It looks like this technology is  applicable to multiple markets. We’ve been   talking with customers from [17:55 inaudible] and  space, all the way to high-performance compute,   and consumer applications. As I mentioned, 3D  IC is a topic that every front-end SOC design   or verification team should make themselves aware  of. The innovations that we see from the market   leaders, such as Apple, will surely ripple down  to all of us. Consumer applications – so, take   a look inside your Apple Watch or your i-device,  look at the new Apple M1 chip family. They all use  

chiplet and die-to-die technology with wide memory  and package, for example. And in larger desktop   devices like the Mac, we see the new M1 Ultra  processor chip just announced last week by Apple,   that uses eight advanced memory chips in the  package for a total of 32 fast memory channels,   that’s 800 gigabytes per second of memory  bandwidth. But more importantly, for 3D IC,   the main SOC in that package consists  of two of Apple’s existing M1 Max dies,   connected edge to edge by a silicon bridge. They  refer to it as an interposer, which hooks up over   10,000 signals that were pinned out along  one edge of the existing M1 Max die to enable   this doubling of processing capacity. That  thing is about 47 millimeters long in total.  

The interconnect alone is 2.5 terabytes  per second of bandwidth from die-to-die,   which is more than four times the bandwidth the  other multi-chip interconnect technology could   provide. And what’s interesting about this  approach is it looks like Apple architected   it that way much earlier in the M1 design process  with the floor planning and interconnect already   layer in the M1 Max chip that was essentially  future-proofing their design. And now they have  

one of the fastest integrated processors and  GPUs on the planet. So, these boundaries are   being pushed by the top of the market, but we  can be sure that the technology and design flows   will trickle down and become more accessible  to all, and that’s part of our job as EDA.   We’re going to see a healthy mix of proprietary  and industry-standard solutions on the menu.   So, we’re seeing a lot of innovation coming from  Apple, Intel, and consortiums, and foundries.   What secret sauce is EDA working on for the  front-end? And what are the flows and solutions   that will grow up with this emerging technology  to help the front-end architect and RTL teams?   Interesting question, John. As we previously  discussed, we offer the ability to capture   alternative designs scenarios, leveraging chiplets  and 3D IC technologies, and the predictive models   and workflows to assess each of those scenarios.  So, as the complexity of these systems increases,  

this can be a daunting task to generate and assess  a multitude of scenarios. So, this challenge   lends itself very nicely to leverage machine  learning technologies, to automate the generation   assessment and optimization of the solutions to  hone in and best meet the design requirements   in a more automated and timely manner. So, I  think this is a key area of innovation that   will extend this technology adoption to a broader  set of customers beyond the current small set of   very advanced users that are in this space today. Absolutely. And we encourage all of our audience  

to invest time in learning this from  day one. And we’re really excited to be   investing to provide you with the tools  and the flows to help you do that.   That’s great. I want to thank you, Gordon and  Tony, again, for another highly informative   discussion on front-end architectural design  verification considerations in this episode   of our 3D IC series. We’re all out of time  today, but we’re looking forward to the next   3D IC podcast with you. Again, thank you both.  And we want to thank all of you, our listeners,  

for listening to our podcast today. Yes, thank you, all. And thank you, John.   Thanks, John, for hosting.  And Gordon, great discussion.

2022-12-19

Show video