Tech ONTAP Podcast Episode 395 – Electronic Design Automation (EDA) in Azure NetApp Files

Tech ONTAP Podcast Episode 395 – How Azure NetApp Files Powers EDA Workloads
===

Justin Parisi: This week on the Tech ONTAP Podcast, we talk about Azure NetApp files and EDA with Andy Chan and Michael Johnson.

Podcast Intro/Outro: [Intro]

Justin Parisi: hello and welcome to the Tech ONTAP Podcast. My name is Justin Parisi. I’m here in the basement of my house and with me today, I have a special guest to talk to us all about EDA and the cloud, specifically Azure. And to start it off, we are going to talk to Michael Johnson from NetApp. So Michael, what do you do here at NetApp and how do we reach you?

Michael Johnson: Hey Justin, thank you so much for having me here on the show today. I am one of NetApp’s Semiconductor EDA specialists. I’ve been at the company for about 10 years, working with all kinds of different semiconductor accounts, as well as partners like Microsoft, looking at how we enable NetApp technologies in the cloud to help customers get better outcomes.

You can reach me easiest ways, probably directly through my NetApp email michael.johnson@netapp.com. You can also find me on LinkedIn just type Michael Johnson NetApp and you’ll find me right up.

Justin Parisi: Also with us today, we have Andy Chan from Microsoft. So Andy, what do you do at Microsoft and how do we reach you?

Andy Chan: Thanks, Justin, for the kind introduction. My name is Andy Chan. I am an Azure global black belt. My focus is on enabling Azure for semiconductor workloads, such as electronic design, automation. Before Microsoft, I have spent pretty much my entire career in the semiconductor space, ranging from design IT to running a entire data center for EDA.

Right now, I lead a Microsoft Semiconductor community that includes folks across all time zones that really provides a high tech engagement model to our customers, to our partners, to our suppliers alike on how to make Azure the most optimized cloud for EDA.

Justin Parisi: Andy, I know that you probably didn’t start doing EDA just at Microsoft. So tell me a little bit about your background, like what you did prior to the current role that gives you the EDA expertise, because that’s not something you just pick up.

Andy Chan: Correct. EDA is a very specialized field. I started my career 20 years ago with a company called Credence that does electronic automated testing equipment. Automatic Testing Equipment is a piece of hardware at the end of a production line to ensure that the chip is being manufactured to spec. After that, I spend quite a few years at Mentor Graphics, a major EDA software provider.

Then I moved into a small startup in Pittsburgh that focused on high performance computing with a high performance caching solution called Avere System that was later acquired by Microsoft. So I was part of that acquisition into Microsoft. Because my background in EDA and in the semiconductor space, I became part of Microsoft Global Black Belt, GBB, a group of specialists within Microsoft that look into a specific workload.

Justin Parisi: And Michael, you also have a pretty extensive background in the semiconductor industry as well.

Michael Johnson: Yeah, I started my career at Philips Semiconductor designing digital TV chips, 2D, 3D graphics, and running a large design verification team. Later went to work at Cadence Design Systems in multiple different roles. And then later Atrenta, which is now a Synopsys company supporting a product called Spyglass before joining NetApp as a semiconductor specialist.

Justin Parisi: EDA is one of those underappreciated industries because it’s so ubiquitous in what we do every day, but people don’t really think about how does this phone get its chips? How does this refrigerator get its chips inside of it? And the semiconductor industry is revolutionizing the way we do these things by making those form factors smaller and smaller and smaller and overcoming these very technical obstacles to try to do that because you’re basically fighting physics at that point.

Andy Chan: Correct. In the semiconductor space, we often used PPA to measure what success looks like, right? PPA, power, performance, and area. Obviously, you want your chip that require less power to operate, so your cell phone’s battery life can be extended.

Justin Parisi: So I guess the area part would be stacking the chips, right? Basically having smaller space for the chips itself, but actually stacking them in smaller form factors, all the way the nanometer level and making these much smaller because they need to fit into these smaller devices, our phones and our Apple watches and that sort of thing.

Andy Chan: That’s correct. And also smaller means the distance between gates is smaller so you have faster response time and you can pack more gates or more logics into a single chip so it allows a single chip to do more.

Michael Johnson: Andy, I would probably add on to that PPA is a key metric for things like how good a battery life you get, like you said timing is how fast it runs and how much performance you get out of it, and a very key metric that we often don’t think about, because the other first two are pretty obvious, but area has a lot to do with the yield.

How many of these little chips can you get on a wafer? And that has a direct correlation to the profitability of that chip for the manufacturers. And so optimizing all three of those key attributes is of huge importance to our customers, our common customers. And that’s something we work very closely with them to help them build more profitable, less power hungry, fast chips.

Andy Chan: Thanks, Michael. And one of the way to improve PPA is additional simulation, running more workload to simulate the behavior of a particular design. and that insatiable demand for simulation. It’s where Azure can really come in to address this very requirement.

Justin Parisi: Yeah, and power goes hand in hand with temperature. Keeping our phones cooler, making sure they don’t get too hot. So these are all very important aspects of the chip design industry.

So did you and Andy know each other before NetApp or is this where you guys first met?

Michael Johnson: I think we met here for the first time.

Justin Parisi: Match made in heaven. Yes. NetApp and Microsoft. So, speaking of that, we have a first party service in Microsoft called Azure NetApp Files and that is storage as a service or volumes as a service. So, Andy, you work pretty closely with that. Tell me about your history with Azure NetApp Files.

Andy Chan: Good question. The first EDA deployment in Azure happened back in 2018, and during that time, Azure did not enjoy a first party service such as ANF so we have to be creative. We have to deploy different file systems in Azure to provide that high performance solution that’s stitched all the compute nodes together to a centralized file system. It was not successful. We were able to get it running but not at the performance that customer was expecting comparing what they already have on prem. So we brought that feedback directly back to our product engineering team.

And basically, position EDA as a key workloads that Azure should focus on. During that time, it happens that Azure NetApp File is coming online. So we were able to quickly match what the customer demands and what Azure was able to offer in a very rapid fashion. We went back to that customer and they are already a NetApp shop.

So the value proposition was a easy thing to make. So we connected them. We start our deployment with ANF. I believe we started with a Spice workload quickly ran into 40,000 cores. And ANF was holding up extremely well. We were extremely happy with the result. Continue to bring that feedback to the product group, and continue to explore ways we can do additional workloads.

Explore ways we can come up with a better reference architecture to really optimize this Fully managed ANF service. How best to run on a particular scale, on a particular tool, if we have a hybrid setting. What I meant by hybrid setting is some customer would prefer to keep all that data on prem only to burst into Azure when they have a surge demand.

Some customer would prefer to run a workload 100 percent in Azure where data lives in Azure 24×7 and they will leverage the full data lifecycle management in Azure to support their EDA runs. So each customer has different requirements. My job is to make those requirements known to our product team, provide active feedback to the product team, to help look into what we can do today with the solution that we have to best match the requirement.

That’s one thing. The second thing is if there’s feature that’s missing or if there’s a wish list that customers want to see in Azure, then we provide that feedback, to influence ANF road map offering. One of the things that people should realize that EDA in Azure is not a final destination.

There’s no stopping point because the environment is evolving. We solve one problem together, people want to do more. There’s a new requirement coming up that people want to leverage them. For example, back in 2018 when we started doing Azure for EDA, People did not ask for AI infusion. Now, most of my conversation with my customers or with my partners is around how to incorporate AI into EDA or to improve chip design. So there’s a lot of work to be done on that front. There’s always new use case to have a new solution we can bring together to address those use cases.

Michael Johnson: I think if I could summarize what Andy said is, over the 20 years that NetApp has been supporting the semiconductor industry, EDA is hard, and the IO profiles and the workloads that EDA runs is really hard on storage, and Microsoft realized that early in their cloud journey and decided to partner with NetApp to bring that tried and true storage platform into Azure as a first party service to make the transition to cloud or the burst to cloud use models easy for customers.

Justin Parisi: Yeah, it’s not unusual to find an EDA customer that’s a NetApp shop because EDA has its roots in NetApp. When NetApp was first started, it was an NFS only storage device, and EDA was one of the very first industries to really leverage NetApp. And the reason why is because It handles those workloads a bit better than some of the other things out there. Those workloads are very naturally metadata intensive, very high IOPS, not a ton of throughput, so you’re not really dealing with a lot of throughput more so than you’re dealing with CPU processes trying to get through all those operations.

Andy Chan: That’s correct, Justin. Because the history of NetApp supporting EDA, the value proposition of Azure with a first party solution such as ANF is an easy one to make to the customer because they are already quite familiar with what NetApp can do.

Justin Parisi: And it’s interesting because when you first implemented the EDA solution for this particular customer, it was on a regular volume. In that case, from ONTAP terminology, a FlexVol, and that’s a single instance of a volume. Now, just recently announced general availability is our large volume. And this is where I think the EDA workloads really start to shine because they handle those metadata operations, particularly the write metadata, the creates, the SETATTRs. Those are done in parallel across multiple volumes underneath the covers as presented as a single namespace. So, I know that you’ve worked a little bit with that. What are you hearing from customers that are doing their EDA operations. Is the scale with a regular volume enough or are they looking to try to scale even more with a large volume?

Andy Chan: Before I answer that let’s look at the difference between Azure and on prem. So on prem, and I’m speaking from someone who used NetApp on prem several years ago. When you make an investment on a high performance storage, you purchase that cluster, or a stamp, and you deploy that, and you stay with that hardware for several years.

And you tend to optimize your workflow within that hardware. You can add additional hardware to it and make your aggregate bigger to get more performance. But that requires certain planning, that requires certain data center considerations. Whereas in Azure, that constraint is no longer there, meaning, when you run, your workload requires certain IOPS, requires certain throughputs, or a certain way to handle your metadata, and you can dynamically increase or decrease the size of your pool in ANF, and you can optimize the way that you like on demand.

That’s not possible compared to on prem with hardware constraints. You mentioned about large volume, and that’s a really good way to look at it, because when a new feature, like large volume, comes online, a customer or even our own first party silicon team can immediately leverage that new feature and look at ways to optimize their flow to leverage that new feature immediately, and if something that did not work out, they can shut it off and restart again without heavy capital investment of purchasing the hardware like on prem. So yes, customers excited about large volume. But another way to look at it is that dynamic allocation that access resource on demand that really take advantage of the new feature immediately, with the goal of productivity gain. Because at the end of the day, you want your engineer to be as productive as possible, and one way to do that is allow them to have access to the resource on demand that allow them to run additional simulation, additional workloads to really push them to look better.

This is where we saw demand really come to being. It’s not the IOPS themselves, it’s how do we translate the IOPS into greater productivity. That matters the most at the end of the day.

Justin Parisi: Well another way to increase productivity is to make it go faster. So, your job’s running faster, right?

Your jobs completing faster because you have more scale to work with. And that’s really where the large volume comes in because now you’re taking this concept of a regular volume, which lives on a hardware. It lives on a single node. within the Azure NetApp Files stack. A large volume lives on many nodes in the Azure NetApp Files stack.

So now you have the advantage of multiple nodes. You have the advantage of more space. You have the advantage of more parallel processing. So now you’re seeing a vast reduction in the amount of time it takes to finish these jobs, which Michael, you’ll talk to us about the value of that, but it really drives the cost down of what you’re trying to do in your EDA space. Michael is an expert in this. So Michael.

Michael Johnson: Yeah, the ideal situation is that you’re compute limited and not storage limited, meaning that you could run as many jobs in parallel as you desire, and the cloud enables that. And you don’t want to be IO limited. That’s kind of the definition of high performance computing.

So, what ANF provides, is the highest performance NFS storage, in any of the clouds today. That enables you to scale more jobs, as Andy said, run more simulations, to do more optimization around AI and reinforcement learning to optimize that power performance scenario. But more importantly, It’s about reducing the overall cost of these exploding semiconductor projects that are very expensive by making engineers more productive, utilizing the semiconductor licenses. Most VPs of engineering say, using cloud actually lowers the risk of a project. It lowers the overall cost of a project. It makes engineers happier. And they innovate better. And that’s well proven. And I think Azure is a great platform for achieving all those goals.

Andy Chan: Thank you, Michael. Another way to look at it is the cost of designing a chip, the largest cost is engineering salary, and the second largest part of the spending is on EDA software license.

And you look at the entire spending, the infrastructure. And high performance storage and compute, it is the smallest part of the pie. And with the cloud, whenever there is a new CPU that’s available, people can immediately incorporate it into their compute farm, which is much more difficult compared to on prem because, folks have to deal with data center considerations.

With newer generation CPU, I have seen anywhere between, 10 to 20 percent IPC improvement. That’s instruction paid cycle. So meaning that with high performance file system coupled with the latest CPU available by Intel, by AMD, or our first party, Cobra 100, folks immediately can upgrade the existing compute farm to incorporate a new piece of technology to have a greater return without like what Michael says, a lot of risk factor into adopting a new technology.

So this is where cloud really shines for EDA because, like I said, EDA is not an end goal, but it’s a journey that’s continually evolving. With new technology, with new methodology. We talk about AI several times, but what can AI do to automate a workflow? And after all electronic design automation, A stands for automation, right?

The more you can automate, the more you can accelerate with newer hardware and newer solutions, then you will make your engineer happier and they will be more productive.

Justin Parisi: Yeah, that’s another aspect of the performance is, the storage can be fast, but if the clients are not fast enough to push enough to the storage to make it work, then you’re not really going to get much benefit out of the workload.

So you could have the fastest storage in the world. And then if your clients aren’t doing enough, it doesn’t matter. So now the clients are faster. They’re using GPUs to try to push this extra work to our storage devices. So now you get to the situation where you might start to reach your threshold of a storage device or a single node.

And that’s again, where the large volume comes in to allow you to scale out further. So as one side grows, the other side has to follow.

Andy Chan: I like the way you put it, Justin, because EDA is not a monolithic workload. It is composed of multiple components, right? You have compute, there is an orchestration layer, and there is the high performance layer.

With ANF being fully managed, that can scale up on demand, people no longer have to spend time worrying about, can my NFS keep up with the demand of the performance improvement of the compute? It is trusted, it can scale up because large volume to the performance level that they need.

So they can spend more time focusing on the orchestration part of the workflow. The compute part of the workflow, if I am adding 20% IPC on my compute side, I can get 20% reduction in my runtime without worrying about can my NFS keep up.

Justin Parisi: Another thing you can do with Azure NetApp files is you can kind of treat it like a cache of sorts.

Right. And lemme kind of expand that. So you’re familiar with caching, with Avere. With a regular volume in Azure NetApp Files, you have data that goes in there and it might be active at any given time for an EDA job. After a certain amount of time, those jobs might go dormant, and if it’s still sitting on this hot tier, this expensive, high performance stuff, you can reduce the tier if you want, or you can enable something called cool tiering, which is basically taking cold data, tiering it off to lower cost Azure Blob.

Are you seeing your EDA customers finding value in that?

Andy Chan: Definitely, there’s a value regarding tiering. Customer has seen value in tiering for sure, because it is part of that access, resource on demand value proposition.

Michael Johnson: About 60 to 80 percent of design data on any given day or week is cold.

And that’s just the nature of the design process. There’s nightly regressions being run, there’s release builds that are built, but then the performance load goes down as you start working on the next builds, because people are constantly checking the code and there’s design changes happening.

When you look at the things like the EDA tools and the libraries, which, are growing very rapidly, that data can be as much as 90 percent cold data. So when you look at putting this stuff in the cloud, and storing the data, you don’t want this cold data sitting on the most expensive, highest performance storage tier out there, Because you’re just wasting money. But it’s difficult to know what should be tiered, when it should be tiered, and ANF just takes care of that and tiers it off to a lower cost tier of storage. But then, let’s compound that. What we’ve seen as designs have moved from 5 nanometer to 3 nanometer, that the amount of storage required per project has increased by four times.

Which means that the cost of storage, if you left it all on a very high performance All flash array is just going to start to explode. So being able to have the ability to tier off and manage that cold data, freeing up space for the very high performance workloads that need the storage, gives you a good cost blend and balance that, I think, is critical, again, for keeping those semiconductor chip development costs down.

Andy Chan: I agree completely, and speaking as an end user of NetApp back in the days having the intelligence of determining which set of data are cold that can migrate off to a lower tier was not possible on prem a few years ago, right?

As technology advance, as solution advance. Having a fully managed service that can automatically determine which set of data can be moved to a lower tier with less cost, it is a huge benefit. Because without worrying about the data set, just focusing on the performance aspect of ANF. If I need to run this workload at that scale with that IOPS requirements a lot of times we run into capacity limitation because my dataset is too big.

And I have to worry about capacity. So, having a ability to tier meaning that I no longer have to worry about capacity. I just focus on the IOP that I need on that particular period of time to support my run. That makes everyone’s job easier.

Justin Parisi: Yeah. And really to stress it, it is automated.

It’s the A in automation. You don’t have to think about it. It does it for you. You’re not worrying about what’s hot and what’s cold. And the best part is when it becomes hot again, it’s still automated. It brings it back for you.

Andy Chan: Yes, you can turn the knob to 11.

Justin Parisi: That’s right.

So as far as cloud trends go with EDA, you mentioned earlier that a lot of EDA customers are looking at cloud for bursting only and keeping their data on prem, and then others are doing 100 percent cloud. So what do you say the percentage split is of 100 percent cloud and burst to cloud only?

Andy Chan: EDA is a workflow that has many stages. For most shops that are doing digital design I would say that 50 percent maybe more of the workflow are on RTL simulations. And that particular workflow fits really well for the cloud because they are less data heavy, meaning they do not have to go back and forth between on prem and in the cloud, and the customer can compartmentalize it 100 percent in Azure it’s all about data gravity. A lot of folks are doing back end as well. It’s all based on the data considerations. Where data is before compute, during compute, and where data is after compute. And they would use that as a consideration on what percentage of the workflow that they would need to put in Azure. Having a reference architecture is very, very useful. We can share with the customer and basically showing them if they want to do a single workload, what kind of data consideration they need to put around that to help them be successful. Obviously, each flow is different.

Our goal is to get them 80 percent there so they can optimize the last 20%.

So this is where working together with NetApp and some of the Lighthouse customers really pays dividends because we can use that example. as a reference. So if I am to do a full chip mixed workflow end to end, what do I need? What do I need in terms of performance? What do I need in terms of full data lifecycle management?

After I have created all those data, like what Michael says, if we are to move that to a core tier, how do I make it, globally redundant? Because most companies have global footprints and certain data sets they need to access to. Well, how about Virtualization,? If I want to look at waveforms, what do I do?

So my way to explain it is not a single approach, because EDA is very complex. It involves multiple components that involve multiple stages, and will involve multiple data protection, data privacy, data gravity considerations. So it’s not a simple answer of what folks are doing, because once people realize the value of accessing resources on demand, they become really creative because you’re dealing with a bunch of engineers who love to think outside the box.

What can I do with A? What can I do with B? And now suddenly I don’t have to worry about my limitation of a high performance file system. What more I could do? So those are the questions that really get me going in the morning, what can I do when now I have all the freedoms of resource access. So that’s a long way around to answer your question. There are a lot of folks doing a single workload. There are a lot of folks offloading certain percentage of the workload that is quite system intense. The idea would be to fully update on prem farm to do, you know, different type of jobs.

There are folks that are already doing full chips, 100 percent in the cloud. For example, there is a start up in the Bay Area. They wanted to do AI chip design a few years ago. They came to us and wanted to work with us. So we put together a team that worked with that company, and after a month of engagement, they were able to run their first RTL simulation 100 percent in Azure. They were doing 7nm TSMC back then. We don’t even have to worry about Investing in the on prem environment. They were born in Azure, 100%. Very, very straightforward. That’s one example. I mentioned there’s a company in Asia that’s doing, you know, SPICE Analog Mixed Signal Simulation at a very large scale.

And they do it on demand. Whenever they have a limitation on their on prem data center, they will burst into Azure and they will tear down the environment when they’re done, so they stop the billing. That’s one good use case. Six months ago there was a paper released by AMD on SCMicro running certain workloads in Azure. This is the beautiful thing about cloud is people are no longer tied to a particular set of constraints that they have to work with on prem. And suddenly, after they moved into Azure, the sky’s the limit.

They became really, really creative. There’s high confidence that Azure is the optimized cloud for chip design. And the software providers are putting forth creative licensing models that are changing how they license EDA software, so really the goal is pushing a greater adoption of cloud for EDA and to further enhance the forward movement. We spoke of large volume, ANF really stepping up in terms of removing the constraint of worry about power systems, performance and limitations.

And of course, the last year and a half, AI is on everybody’s radar. People are asking what we can do together as an ecosystem with advanced AI. How do we leverage LLM? How do we use our RAG to improve the chip design process? Microsoft has put out several videos and have given several talks about That last year’s DEC – Design Automation Conference, for example, how we leverage ANF, how we leverage AI for our own chip design.

So I do see a great synergy out there with the ecosystem. And what I mean by that is, about 10-15 years ago when I worked in this area, Folks tend to be operating in silos because you want to keep your flow a secret, right? If everybody has access to the same resource, the guy or the gal that can optimize your EDA workflow better, maybe 5 percent faster, then you’ll enjoy that 5 percent performance gain over your competitor.

So they tend to work in silos. There’s not a lot of conversation between companies. That have changed a lot because with cloud, people realize that there is win win situations out there by sharing best practice. One must be extremely respectful of their IP or the certain things that should not be shared.

However, there is a lot of items that the ecosystem or the vertical are sharing and folks are a lot more open what they can do together to create a win win. There’s a lot of industry events out there that really focus on cloud, that win win proposition. You know, you look at TSMC’s OIP or Symposium.

You look at industry event like User to User, Cadence Live, Synopsys events. They really push the idea of how can we work closer together. So Michael, what’s your take on that?

Michael Johnson: What I see in the industry is that the largest of the semiconductors seem to have an economy of scale that enables them to maintain huge data centers. And they also have legacy data centers that they don’t seem to be adopting cloud as fast. But where we see very huge adoption are the smaller semiconductor companies that are realizing that their next project is going to need four to six times more compute and four times more storage.

And they’re already in a full data center. They’re at an inflection point where it makes sense for them to look at moving partially or fully into cloud. But then there’s also a bunch of companies out there that I am fascinated by that are new to semiconductors. I was recently on a call with a financial trading company that was doing EDA work.

And I’m like, what? Well, apparently they’re building some sort of device to help them, I presume, do financial trading faster. That’s unique to them. And I’m like, well, okay, well, we can help you with that. But they now need a unique environment and they can either build it or get it from the cloud.

So I think those types of examples are there. I think the other one that Andy touched on is this whole AI thing. When I talk to the three EDA vendors, it’s clear that AI is as important to them as it is to NetApp and to Microsoft, and we’re all in, and just about all of the EDA tools are becoming GPU enabled, at least the ones that particularly strongly benefit from it.

So now, your IT team needs to say, well, what percentage of servers in our high performance computing farm need GPUs? And all of a sudden, they’re going to have to start reconfiguring in their environment. If they move to Microsoft, they just say, Hey, I just realized that my workload benefits from enabling the GPU capability of this application.

And guess what? Microsoft can provide those services. So I think you’re seeing a lot of dynamics change in the way that companies are looking at the scaling compute environment and how to address it.

Andy Chan: Another way to look at it there is a core EDA component and people do a shift left. They take the existing environment and they migrate it off to Azure at the speed, at the scale that makes sense to them.

I think that is firmly in place and as most of the EDA companies or even startups that are doing a particular customer silicon design? What we’re seeing today what Michael was pointing to is smart people already thinking, right? After I migrated my EDA workload to Azure, what else we can do using EDA as a core?

What else I can supplement to my core, like machine learning, the new AI capacity that’s running as a service in Azure. What can I do with those? We speak of data having gravity. What else I can do with my data set? Can I replicate my data to a different geographic location that allow the folks that are doing the front end, doing remote desktop, allow them to be more productive, right? Can I shut down a data center in Asia that is not well maintained and move everyone into a Azure data center nearby? We moved the need to maintain all those small data centers and this is an actual example of a EDA customer who had a data center in India that’s not well maintained.

So after they moved into Azure, they were able to remove the need to have a local data center and just use Azure backbone to improve data replication, and suddenly sky’s the limit, like I mentioned before. Now, there’s a different virtue of going to Azure take BCDR for example. Almost all the EDA companies or chip design shops has to have a BCDR consideration.

One of the tangible benefits after they migrated a portion or some percentage of the workload into Azure, suddenly their data is being protected, and they have a dynamic of moving from one Azure data to another one for their batch workload. So they’re no longer tied to a single region. And they realize that they might have a BCDR solution right there. So all those intangible benefits that people can see where EDA as a core expanded to the edge with different use cases and different benefits. And I think the conversation is moving from just shift into the art of possibility.

Justin Parisi: So if you had a list of advice for EDA customers or EDA workloads that are looking to move to cloud, what would you give them to help them start? Say top three things that you would have them do before they got into the cloud?

Andy Chan: My vote is to use the 80 20 rule. Don’t try to make things difficult. If you do that, you get frustrated and you give up. So do the low hanging fruit. And what are the low hanging fruits? My vote is RTF simulation on IP or block level. They tend to be small, ranging for 4 to 8 gig per core per job. And the data dependency is not that heavy, so you reduce that data gravity consideration.

And, like I mentioned, for digital design 50 to 60 percent of the workloads are RTL simulations. Second workload are your analog mix signals, and they tend to be small as well, and while some of the tools might not scale out as RTL simulation the data dependency is not as bad as some of the backend workloads where the GDS2 file can get pretty big.

So that is the second workload that I would vote for. Thirdly, is the digital sign off, STA. You can use your IRDrop or your DRC as a single bundle. Folks tend to put in that free category. So RTF simulation is carry one, pretty straightforward, move to the cloud, analog mixed signal is the second group, and then digital sign off as the third group. Each company has different considerations there are different requirements and why I love Azure for EDA is we put together a dedicated team there are different team members at different regions from EMEA, from Asia, we assemble them and whenever there’s a company or there’s a group of folks want to migrate the EDA into Azure, the same team that I mentioned would be fully engaged. We provide a high touch engagement model where we lockstep with customers to give them direct access to the subject matter experts, to the product groups, give them direct access working with NetApp, for example. To our ANF team and NetApp has a dedicated team of folks that enables EDA in Azure. All those high touch engagement really reduce the people overhead of leveraging cloud for EDA.

And at the end of the day, people matters. And you want to provide that personal touch to really make the project successful.

Michael Johnson: Yeah, Andy I would concur with your list. I think for the podcast listeners, one of the questions I get is, does that tool work in the cloud?

If you’d asked me four years ago, I might say, maybe, I think so. At this point, I would say a definitive yes. All three of the major EDA vendors have run full workflows in the cloud for years. So if you’re going to start a cloud journey, Reach out to people like Andy, there are well tested reference architectures and we can tell you how to do it successfully so you don’t have to learn on your own. The second thing I would say clarifying Andy’s point is he mentioned RTL verification workloads.

I would go further and say there are a whole bunch of jobs in the EDA workflow that run every night. Nightly regressions, code check ins, chip finishing jobs that run night after night after night, or for many nights, In a fully automated fashion, RTL verification can take up as much as 60 percent of the compute in your data center.

Take those big, recurring jobs, move them to the cloud, let them spin up when they need to, let them run, frees up a massive amount of resources in your data center, but has very little impact because those workflows are typically owned by a small group of your automation specialists, the people that are doing your build and release processes and such, so that you don’t have to go through that learning curve of teaching all your engineershow to use the cloud and then as you as a company build up that skill set and that maturity then you can start onboarding engineers who will probably find that they like the cloud as much as on prem or more.

Andy Chan: The goal here for Azure is to create. an engagement model that is faster, more convenient, and to do that we put together a team of subject matter experts. So I kind of repeat myself here, but this is very important because migrating EDA into Azure is not a simple task. Let’s be really upfront about this and require people. This is where I would love to work with, our potential partners here in how do we do this from a people perspective. And this is the only way to do it. I cannot move a single workload if I don’t have the people and process in place.

Justin Parisi: All right, Andy, Michael, thanks so much for joining us today and talking to us all about EDA in Azure. So, Andy, if we wanted to find more information about this, where would we do that?

Andy Chan: Yes, you can reach out to me directly, andy.chan@Microsoft.com. Additionally, we have several white papers and blog posts. You can find it at community.microsoft.com. And thank you for inviting me to speak with you today, Justin. It is a fun experience. I love to talk about EDA.

Justin Parisi: Yeah, thanks for joining us. And Michael, do you have any where we can find information?

Michael Johnson: Yeah, I think a quick Google search of Microsoft NetApp EDA will hit on a whole bunch of things, you’ll hit on the reference architectures that Andy referred to you’ll see the people that have been writing blogs and communicating on this, and of course, reach out to Andy and I, we can easily connect you to the right people and get you started right away.

Justin Parisi: And of course, there’s learn.microsoft.com, where we have all of our Azure NetApp Files, documentation, some of that written by yours truly. So thanks again, Andy and Michael.

All right, that music tells me it’s time to go. If you’d like to get in touch with us, send us an email to podcast@netapp.com or send us a tweet @NetApp. As always, if you’d like to subscribe, find us on iTunes, Spotify, Google Play,iHeartRadio, SoundCloud, Stitcher, or via techontappodcast.com. If you liked the show today, leave us a review. On behalf of the entire Tech ONTAP podcast team, I’d like to thank Andy Chan and Michael Johnson for joining us today. As always, thanks for listening.