Intel Corporation (INTC) Earnings Call Transcript & Summary

All right. Good morning. Welcome, everyone. My name is Michael McNerney. I'm the VP of Marketing and Network Security here at Supermicro. And this is Supermicro's CTO Panel on Innovation in Cloud Infrastructure. Today, we are very pleased to bring together thought leaders from across both cloud hardware and cloud software to discuss the latest innovations in the open cloud infrastructure. For those of you unfamiliar with Supermicro, we are ranked as IDC -- by IDC as one of the leading global providers of server and storage systems. We are based -- we are U.S.-based with our headquarters, operations and primary R&D located here in San Jose, California. And no, we are not moving to Texas. For our panel today, we wanted to specifically discuss the interplay of open -- optimized hardware in open software-defined solutions and how that can bring value to the enterprise. We see this value really showing up in 3 ways. First, application optimized systems, systems that are designed and optimized for a specific workload. We've done a lot of work with both VMware and Red Hat to build reference designs that match a given software solution, vSAT running on a 1U Petascale NVMe system that can deliver 10x the performance improvements. A Red Hat OpenShift optimized on BigTwin multi-node systems to improve efficiency. Matching the software requirements to exactly the best hardware configuration delivers performance, cost and efficiency. At scale, these efficiencies deliver tremendous value. Second is the importance of first-to-market innovation to deliver new services and competitive advantage. Innovation at the hardware level can unleash entirely new applications, solutions and markets. The disaggregated SuperBlade systems that we collaborated with Intel on deliver up to 60% savings in acquisitions costs. I will let Shesha talk about that more, but last time I checked, Shesha had over 250,000 disaggregated systems, delivering significant first-to-market advantage to his business. So without further delay, let me introduce the CTO panel. First, Chris Wright. Chris Wright is the SVP and CTO at Red Hat. Chris is responsible for incubating emerging technologies and developing forward-looking perspectives on AI, cloud and edge computing. A software engineer at 25 years and a Linux developer working deep in the Linux kernel for much of that time, he has risen through the ranks at Red Hat to become an authority on the merits of taking a community-powered approach. Kit Colbert is the VP and CTO of Cloud Platform BU at VMware. Today, Kit is tasked with delivering technical strategy and innovation for VMware Cloud Foundation, vSphere and VM's hybrid cloud offering. An avid VMware evangelist, you'll often find him on the main stage of VM world in giving valuable strategic advice as the hybrid cloud voice of VMware Office of the CTO blog. Shesha Krishnapura is a Fellow and IT CTO at Intel. Shesha is responsible for advancing Intel data centers for energy and rack space efficiency, HPC and optimized platforms for enterprise computing. Fostering unified technical governments across IT, leading consolidated IT, strategic research and pathfinding efforts and advancing the talent pool within the IT technical community, all fall within the remit of this tech visionary. A 3-time recipient of Intel achievement awards, Shesha was appointed an Intel Fellow in 2016, and his creativity and leadership have been recognized by U.S. Department of Energy for industry leadership in energy efficiency. And finally, Vik Malyala. He's an SVP here at Supermicro. Vik manages worldwide field engineering, solutions and technology enablement teams, responsible for understanding the largest enterprise and service provider customer requirements and deploying optimized solution to meet those needs. Before we begin, I'd like to highlight the questions tab below the screen. Type your questions into the box provided. These will be posted to our guests in a live audience Q&A at the end of discussion. So please stay to the end to get your questions answered.

So let's get this kicked off. Let me start first with Chris. Do you -- we talk a lot about cloud and open architectures here today in some of the preamble, do you see that open architecture slowing down or really speeding up innovation in the enterprise?

Well, I love the question because there's a lot of nuance in there. So let me kind of give you a quick view of how I see open architectures. Obviously, I've spent a lot of time in open source, so open source for me is going to be a critical part of describing an open architecture, but it's not the only piece. I mean open architecture includes interoperability, API definitions, open ecosystems, et cetera. So it's not just about open source. And if you look at history, I think you can say that there's an interesting relationship between standardization in the formal sense, which is standards bodies, SDOs, and standardization in more of the kind of industry best practices sense. And I'd say, at this point, the velocity of community development in an open source community well outpaces standardization. So in that sense, a traditional standardization process for, say, a well-defined open architecture definitely slows things down. Even when you look at open source communities and the high-velocity development environment that comes in open source communities, there is sort of a ramp-up time in getting things -- getting consensus and getting things moving and kind of getting over that inertial barrier. In the end, having a common view of what technology investments look like and what platforms look like across not any one industry vertical, but all the industry verticals is where I think we see real sort of speed up across the industry. So if you look at a technology that's well understood, like Linux, that has created a focal point across the industry, I'm not talking about Red Hat and Red Hat Linux, I mean, just broadly, the Linux project, it's created a focal point across the industry. And with that focal point, you do get speed. So I think there's an interesting balance of -- it takes some time to agree on what is the open architecture. But with that agreement, you get real velocity because now we have cross industry focus on a common either best practice, best architecture or even specific software projects like Linux.

Fantastic. And just to clarify, I think we've got Shesha tagged as James Thornton. So that's actually Shesha Krishnapura. Thanks Shesha. Shesha, maybe do you have anything to add there in terms of when you're building out your sort of large scale data centers, how does that sort of open architecture play into it? You're muted, Shesha.

Thanks for letting me know. So good morning, everybody. And how open architecture has helped Intel? So it takes a little bit quick history here. I joined Intel in '91 in my 30th year and the first 10 years in the product development. And it came to 2001 to actually drive the entire EDA, electronic design automation industry from RISC/proprietary UNIX to x86 Linux environment for the entire industry, not just for Intel. So since I joined Intel in '91, we were all using RISC platforms to design [ x86 ] processors. When I joined, it was -- Pentium was just being designed [indiscernible] and all that. So the question was we always wanted to use x86 architecture to design our own next-generation chip. So that took around 10 years for us to get there and the Linux actually helped us to get there with a very open architecture. In the early '90s, it was either HP PA-RISC, HP-UX or IBM RS/6000s with AIX or with DEC Ultrix or SPARC Solaris. So there were various different choices we needed to do. So when we drove this Linux into EDA industry, thus, immediately around $35,000, $40,000 server became a $10,000 server and you get 3 to 4x faster performance and things completely changed. So the Linux paved the way for us, both Intel and AMD in the x86 and x86-64 world for the world to explore in the data center area. So it has helped, the Linux has helped greatly. And then the open architecture -- coming back to what we want to use. So people have a choice. If Intel is a much better platform, then people will choose Intel. If there is a competitive platform is better, they choose. So there is an interoperating play. And that puts both companies to compete for the customer and also get a much better performance, innovate faster. Same thing with the Linux, okay? There are multiple distribution choices, and people can go where they can get much better support and service. So open architecture has evolved where my capital, which used to be around $200-plus million in 2000, year 2000, and now I'm still a little bit above that, but the amount of compute capacity and the scale I get is probably around 50x more. So that's what the open architecture has helped. It has made an economic sense, and it has made technological advances to get a much better performance at a much lower power and hyperscale and so on and so forth.

Yes. So maybe -- so Vik, maybe here, Shesha talked about sort of this adoption of x86 hardware sort of just collapsing the cost point and bringing a ton of innovation. What's next? What's out there on the horizon in terms of hardware innovation that's going to sort of rock the enterprise?

I think, ultimately, it's a customer-driven thing, right? So we looked at it when you talk about the adoption of open architecture. Like, say, OpenStack is a classic example where it was adding a lot of features in terms of provisioning hardware and making it easy for customers to adopt, but enterprise customers just couldn't take it unless somebody like VMware comes into the picture and Red Hat comes into the picture to bring those features into something that can be supported for customers to adopt it because no one want to risk their job in taking some latest and greatest into the IT infrastructure, putting the company at risk. So that's one thing where this kind of adoption of latest technologies will actually help in the cloud adoption somewhere. So what I have seen talking to many customers is that we are at the best time of the IT industry in terms of how many things are coming, which requires a huge IT infrastructure and a definite need for people to manage it effortlessly and be able to scale. And most importantly, what's happening in the last few years, I mean, Shesha can attest to that, power is a biggest, biggest headache for any IT infrastructure. Especially, as more and more systems getting added, the data centers are getting bigger, things are moving to the edge, in any way, any how you look at it, the amount of compute needed is growing enormously, the storage and networking technologies that are catching up with it to keep this compute fully used. But at the same time, hardware architecture point of view, we have to bring something that actually meets the customer requirements and make it easy to scale, make it easy to deploy, easy to manage, and it should not cost an arm and a leg, which also drives for standardization. x86 is a classic case. Why people use it? Because it's there, it's ubiquitous, and we can use it as an open architecture. And with the likes of all the software that's coming in that we can easily deploy on that, customers can adopt it. So from Supermicro's point of view, when I look at it, customers basically are looking for ways to upgrade the compute, storage and networking elements of their data center independently because all of them are moving at a different scale. If you take a look at Intel, every year, they are churning out new processors with more cores, faster performance and more I/O and all that. But if you take a look at and implement in a data center, they're not going to change their network just like that. It's going to move it at a different time. But then we don't need to necessarily throw away the equipment and do that, and that's part of the reason why collaboration with industry leaders like Intel comes into picture. And also, in order to use these things, we want these things to be taken easily into the data center, which in other words mean that qualifying the hardware, whether it's with VMware, whether it is with Red Hat. The main reason why we do that is for people to take it, no questions asked, they just plug in, load the software and get it running.

So let me jump in there. I mean I think, we see this -- on the hardware side, we've got all the stuff coming in, memory, storage, compute. On the software side, it's hard, right? Because you've got to take this new technology and try to optimize software for it, optimize compilers, there's a lot of technology. So Chris, how do you sort of look at adopting these new hardware technologies and bringing them into your stack?

Well, first, we think in terms of -- from a customer point of view, so that -- there's a disruptive change at the hardware level. The beauty of a common infrastructure and common API or open architecture we started with is creating consistency for customers. So where possible -- this doesn't always work. Where possible, we want to -- software is about creating layers of attraction, right? So how do we create the right layers of attraction so that customers are taking advantage of advances in hardware, where despite the change, the CPU speed of a single thread is increasing, but not at the same rate it has been? The way you're getting system-level throughput is system-level changes. So there's more cores, but there's also accelerators. There's infrastructure within a single server that's changing the connectivity between I/O devices and CPUs and memory. So that whole underlying system's infrastructure is changing. We're trying to preserve as much as possible the common experience above that change in infrastructure, again, where it makes sense, and also recognizing that we're entering the space of highly heterogeneous compute. And heterogeneity, while it allows for application-specific accelerations, it creates a level of complexity. So balancing that complexity with automation and management is the trick. That's what everybody is trying to strive for. And I think we've seen a lot of advances, starting in public cloud, showing some different ways of thinking about compute. And I think initially, public clouds had this notion that all compute was created equal. And well, okay, a little bit of small, medium, large T-shirt sizing to fit your workload, but really, it's about compute equals compute equals compute. Today, if you go to a public cloud, you'll see well over 100 different machine types or flavors or instance types, which just -- and those are all hardware optimizations focused on specific application type of workloads. And so we see that similar behavior happening in the on-premises data center. And so how do you create that cloud-like experience, easy access to a set of acceleration technologies to improve your utilization of power and space to produce the business results that you're looking for? That's a big problem.

So let me jump in there a little bit as sort of a hardware geek, if you will. Like -- so Chris, from your perspective, what's the sort of upcoming hardware technologies that you think you're going to have, some of the infrastructure upcoming that you think is really going to bring some great value to your customers?

Well, I think you got to start looking at what are the workload changes. So traditionally, we focus quite a bit around applications. And those don't go away, right? Business logic is what's building businesses. But increasingly, data has become an essential part of all the discussions. And so ways to accelerate the movement, access to, storage of and processing of data are the kind of underlying hardware techniques. So in the Intel world, you see Optane technologies, you see different types of AI accelerations, and you see different power envelopes for, say, edge compute, where you're bringing compute closer to the actual data. So those are sort of the hardware level changes. And if you look at network interface card, it is no longer just feeds and speeds, right? You're getting a lot of offloads right in the card, including cryptographic offloads. And so connecting data, moving data, storing data and processing data, all of that becomes a central part of the application end result for the business value that you're trying to generate as a business.

And Kit, how do you see this sort of playing out as well? How do you see that hardware innovation coming to play?

Yes. I mean I'd echo a lot of what Chris just talked about, seeing really an explosion of innovation at the hardware level and a lot of heterogeneity as he talked about as well. And so one of the areas -- so I think a couple of things. So first of all, greater applications, a lot of the stuff is there as we talked about for the apps to support these newer apps that are emerging, but it's also some really interesting opportunities from a cloud infrastructure point of view. And so one of the technologies I'm kind of most excited about from that lens is around the SmartNIC or the DPU, the data processing unit, as NVIDIA calls it. And the reason for that is a few fold. So the notion about the SmartNIC is that you now have a general-purpose CPU on the NIC. It's actually a whole -- it's like a little Google server essentially on the NIC, on the card. And that's really cool because it gives you this additional control point that now you can actually have multiple instances of an operating system running on the hardware, one on the x86 side and one on the NIC. And so what we can do then is start using that NIC as a control point for the server and start doing some really, really interesting capabilities with it. So things like what AWS has done with their Nitro architecture, like really rethinking the Hypervisor. We're doing something similar. We call it Project Monterey, announced it a couple of months ago. But the idea there is that it really gives you a lot more flexibility in terms of how you architect both the server and the cluster and can allow you to manage a lot of that complexity that Chris mentioned in terms of all of these new hardware components, accelerators, et cetera, and control and simplify how you deliver those to the applications that need them.

Yes. So all these technologies are great. And so Shesha is sort of the implementer of these, right? It is like everyone's got new toys. How do you implement these at scale? How do you sort of look at adopting? When is the right time to bring these things in? How do you approach that?

Yes, it's a great question. So since you're talking about scale, so let me first talk about the scale. So we have 17 data center sites with 56 modules. We have 85-megawatt worth of IT load data center. We are a hyperscale data center. We are not just an HPC. And I'm just looking at it. In September, we had more than 316,035 servers, just to make it an exact number here. And more than 490 petabytes of raw storage and more than 621,000 network ports. And out of that 85 megawatt, we are running 36 megawatt worth of data center at 1.06 PUE. And our each rack takes 43 kilowatts and our racks are 20 inches wide, 60U tall. And the reason it's 20 inches, not 24 inches is that you can actually put a lot more servers because any IT equipment, server and storage network, since it's an [ open stand ], it's all 19 inches wide. There is no reason for you to have a 24-inch rack. So now going back to the story of what makes sense into the technology. So first, obviously, you need to understand what kind of workloads you're serving. Inside Intel, for example, we have designed workloads where -- all the electronic design automation workloads, where 94% of that server capacity, what I talked about, pushes that design. The next is the traditional IT workloads, like Office and Enterprise, whether you have SAP workloads or any kind of web services and so on and so forth. Then the manufacturing computing, because Intel is also a manufacturer on the silicon part of it, so we have another 3% of the servers there. So 3% in Office and Enterprise, manufacturing 94% in the design computing. So the question is that how do we bring the technology is, first, you need to understand your workload, what are the workloads. We have several thousands of applications and each applications have a different behavior. And if you look at the jobs, we run around 212 million [ bad ] jobs per week for the design. And if you look at the jobs, the amount of memory required for each job is all over the map. A good chunk is within 4-gig RAM and then 4 to 8 gig RAM, and finally, it's around 3 terabytes plus. So -- 6 terabytes. So the question is that how much is the quantity you need, what kind of workloads you need, and first, you get those as a benchmark. Then you look at what's the best performance you can get, best throughput you can get and then what the cost of owning that and then what kind of reliability and scale you can get. So then the operating cost comes into play that how can I reduce the number of cable -- how can I reduce the number of power cables and network cables? How can I get more redundancy with less number of units? And how do I scale? So there are many aspects of it, it comes. It's not just you go into an element of a chip or an element of a driver, but you need to look at the holistic approach about how do you run this hyperscale data center at extremely economic sense, highly reliably and highly scalable stuff. That's what it all comes into play at the end.

Okay. Great. So hey, I know we're sort of getting -- time is running fast here. So I think the one question I wanted to come back to Kit on is just the competitive advantage, the sustainability of sort of running and operating your own cloud. I mean where do you -- when customers come to you and say, why am I running my own cloud, what do you tell them?

Well, usually, the way it goes, we say, hey, what are your use cases? What are the apps that you have? And then we sort of take it from there. And certainly, we see a lot of people go into the public cloud. But we do actually see a lot of reasons to stay on-premises either in a data center or perhaps an edge location like in a retail store, let's say. And there's many aspects of that, some the security or compliance angles, other times, there's various sorts of performance specifics they want to get into. Sometimes, public clouds have support or have the sort of technologies they're looking for. Sometimes like the latest and greatest GPU type or other sort of accelerator may not be available. And so it's really kind of those factors that drive it a lot. And what we want to provide is, really, that sort of choice, right? You can go on-prem or in the cloud and that's perfectly fine. But let's see, I kind of lost my train of thoughts. Did I answer the question? Or...

Sure. No, no, that's great. I know this is sort of a general question. It's great to have this sort of audience on here.

If I may, Kit, I agree with all your points. But the one point which you just missed is the cost element, okay?

Yes.

So I just want to make sure that the cost is extremely important part. Everything else is perfect.

I knew [indiscernible]. So I think cost is an interesting one, right? And I think people tend to ask, okay, is one thing cheaper or more expensive? And I think it's really highly dependent on the workload and the application. Cloud is very good for certain types of things, especially things you need to scale up really quickly, things that are bursty in nature. If you do understand the workload, then you kind of understand the trend of it. And you have a certain average or median usage that you can size for. You can actually dramatically reduce your cost going on-prem. So I do think cost management is another important consideration there. Thank you for reminding me on that one. It is an important factor.

Shesha, maybe you could dive a little bit on cost in terms of building out your...

Yes. So see, the reason is that it's not -- the question is that what's the best way of serving your workload and level of performance you expect -- experience you are expecting for the users and at the lowest possible cost, okay? So those are all the things, if it is business objectives, which you need to take into consideration as a rule of thumb. If you are -- in a year, if your utilization is touching 50% or higher, always on-prem will be cheaper, okay? That's given. The second thing is that, okay, if your workload is -- you need to bring back -- move the data, bring back the data on a continuous basis, then -- and also that has a talk for bringing back to the data, so you need to consider those costs as well. Then that means that if you need to bring back data and you need to have on-prem certain amount of capacity, then that cost with the cloud, public cloud versus private cloud cost, you need to really look at those things. But I agree with Kit that -- the answer is that where you -- if you are a peaky of workload, it happens once in a time, and those kind of workloads you want to identify and see what makes sense to you and how it will help you in managing your overall cost of ownership. So that's where I would probably say that how the cloud picture will come into play. I don't want to repeat all the points which Kit -- which I agree with all the other points.

There's a lot to digest here. Hey, I wanted to jump to a few questions. I really appreciate everyone's valuable time here online as well as our panelists. They had a few questions. One was sort of the role of new storage technologies in the cloud like persistent memory, NVMe. Maybe, Vik, if you wanted to sort of talk to that real quick.

Yes. Sure. I think one of the things that we have seen a lot more interest from the customers is to see how they can take care of a bigger memory footprint because you're talking about more number of cores and accelerators and all these things coming in, which can access and make use of bigger main memory footprint. One of the things is the DC persistent memory. So at this point, what I have seen is very few applications are able to take advantage of it, but there is a significant research is being done by many companies trying to take advantage of it. The easiest way to adopt this persistent memory, we also have seen is the Optane drives per se, right? I think one of the things that we have seen is in a VMware environment, especially like a vSAT environment, you can actually get pretty good benefit from that, and it's already proven and which accounts for like more than 50% of the deployments in a cloud. So the way I look at it is both these things are going to make a significant impact going forward, mainly because of the lower latency and a higher memory that is available for the applications to take advantage of. And as I said, again, the GPUs and accelerators that are taking place in these -- finding their place in these systems, this is going to be, I think, one of the key things to look forward to in the next few years.

Yes. If I can jump in there real quick because we actually have been looking at persistent memory in particular. And I think there's a couple of different operating loads that you can do, right? You could just use it as a normal drive or [indiscernible], it's all the same. And you do get significant -- or good performance boost there, but where the real craziness comes in is if you actually update the application logic to use persistent memory more natively. We actually tested this out. We made a new version of Go or an extended version of Go that could have these sort of language primitives to support PMEM. And then we rewrote Redis in Go using that, and we found something like -- it was like 5x faster, insane performance boost, not to mention insane start-up speed as well. So I think the opportunity out there for persistent memory is absolutely enormous, and I think we're really just barely scratching the surface today.

Yes. I think we see that, I mean, when we look at like the hardware innovation, when you look at cores, memory performance, storage performance, networking performance, I mean, we're always talking 10x. And it sort of comes up and bubbles up to the software level. It turns into like, oh, we got a 20% benefit. Like, hang on, how did 10x turn into 20%? And so how do you see that? And maybe I'll come back to Chris. How do you see that sort of hardware benefit? What's the key to sort of getting it exposed at the software level?

Well, there's a complexity question in there. So for example, the more you're exposing primitives to applications or the applications the way Kit was describing actually need to change, the more challenging it is to get those to sort of percolate throughout the data center, right? I mean the notion that it's -- innovation is turning out change more rapidly than an enterprise can take it on is a very real, like, practical constraint. And so the one trick is make sure your systems infrastructure is not exposing fundamentally new primitives to applications so that the applications don't need to change at all or substantially in order to take advantage of those. And in some cases, that can be done entirely through the low level enablement of the operating system, systems libraries, the things that are dependencies for your application. So maybe you have to relink the application, but you don't have to change the application or recompile it. When you get into application level changes, it's just more work. That's all there to it. And that -- then you have to look at the cost benefit analysis. And there, you do look for what are the 10x multipliers because if you're just chasing more percentage points off of your performance, the cost of the change doesn't necessarily justify the benefit. And so systems -- I think it's the systems architecture view that's really what's changing here. And you see micro benchmarks can always be amazing. The challenge is to make those micro benchmarks meaningful in an application, which doesn't sit there in a tight loop doing whatever routine that you're micro benchmarking.

Okay. We got a few questions more on the sort of DPU space. And maybe, Kit, you could come back, and you talked a little bit about this. Where are you seeing sort of the DPU technology having its first impacts?

Yes. Yes. So I think right now we're seeing it mostly for networking-specific use cases. So we look at like the telco space or large enterprises. And as you move from 40 G to 100 G like -- and above, what you're seeing is that it's becoming more and more difficult to saturate those links and you need to throw more and more [ DPU ] at it. And this becomes more and more expensive. So it's not really working for a lot of these customers and sort of bleeding edge requirements. And so that's, I think, the first kind of killer use case for the DPU. And then as we talked about -- I think there's a lot of other ones that come on after it. So there's things like a lot of storage acceleration, maybe using that as kind of the I/O hub for all sorts of data, both network and storage-related. There's things around supporting this new Hypervisor architecture that I talked about, which can even support bare metal, right? Kind of this crazy kind of great way for cloud service providers or people looking at doing multi-tenancy can come into there. There's a whole sort of security angle that we talked to many customers about as well. So I think, again, we're kind of at the tip of the iceberg here, but the kind of killer use case today that will bring these things into the enterprise, into telcos is focus on network performance. And then once they're there, we can start leveraging them for all these other really cool use cases as well.

And I'm also seeing things like these things are being used for -- to your point, like, security is one of the things where people are taking the route of trust and kind of extending it on to these DPUs. And another thing I have also seen is customers looking at expanding these accelerators or GPUs within a system across multiple systems within a rack while keeping the latency low and being able to move the traffic in a controlled manner. So those are the type of things is what I see that immediately people are looking for. But the -- again, everything gets back to how people are trying to port their applications and how they're trying to move the data because moving the data is expensive. And how they can make it easy for people to kind of manage it, I think that's where, I think, these DPUs are going to come into picture. So the processors and GPUs can do what they do best, offloading all the other stuff onto these DPUs. So definitely, a good market I see for that, and there's a benefit that people at scale can use that.

Fantastic. Unfortunately, that is all the time I have. I do sort of want to thank everyone who joined today. I really want to thank the panelists. We're sort of honored to have such a great set of people, Kit, Chris, Shesha and Vik here. Any questions that we didn't answer we'll answer offline. And you can -- for related content, you can go to supermicro.com/cloud. There's a lot more material that we've been putting together around the cloud things. And we value your feedback, so please go ahead and enter your feedback at the end of this. And again, thank you Kit, Chris, Shesha and Vik for joining us, and we'll, hopefully, talk to you soon. Thank you.

Thanks, Michael.

Thank you. Cheers.

Thank you.

Thank you.