Marvell Technology, Inc. (MRVL) Earnings Call Transcript & Summary

Welcome, and thank you for standing by. I would like to inform all participants that this webinar is being recorded and will be publicly available. Parts of this conference can also be reproduced in JPMorgan Research. If you have any objections, you may disconnect at this time. This communication is provided for information purposes only. I would now like to turn the call over to Harlan Sur.

Great. Thank you, and good morning, and welcome to JPMorgan's hosted virtual webinar with the team at Marvell. My name is Harlan Sur. I'm the semiconductor and semiconductor capital equipment analyst here for the firm. So as many of you know, we've written quite extensively on Marvell's leadership in networking and optical connectivity within the cloud and hyperscale data center markets and their growing market position within the custom chip or ASIC market, also targeted at cloud and hyperscale titans, as these customers look to differentiate their compute, their networking and their storage architectures. And in fact, I think Marvell is really the only semiconductor company that is focused on cloud and hyperscale merchant and custom solutions that focuses on the foundational data center building blocks of compute, networking and storage. And this is really relevant for today's discussion, because next-generation data center architects are focusing more and more on driving continued performance to address increasing complexity of compute workloads, by more efficiently scaling compute as well as storage and memory access and, of course, driving for improving economics or total cost of ownership. So in other words, how can compute have more access to memory to unlock more performance? How can CPUs, DPUs, GPUs and ASIC accelerators not only communicate more efficiently but also share their associated memory and also take advantage of new memory architectures like persistent memory? So this is what the new CXL connectivity and interface technology will enable and the Marvell team will address the CXL standard today, their efforts in implementing CXL into silicon and the time line and market opportunity for CXL. So with that, let me turn it over to Ashish Saran, Senior Vice President of Investor Relations at Marvell to introduce the rest of the team.

Good morning, everyone, and a big thanks to Harlan and the JPMorgan team for hosting us today. Let me now go over our safe harbor statement. Except for statements of historical fact, this presentation contains forward-looking statements, including, but not limited to, statements related to market trends and to the company's business and operations, business opportunities, growth strategy and expectations and financial targets and plans that involve risks and uncertainties. These statements are not guarantees of results and should not be considered as an indication of future activity or future performance. Actual events or results may differ materially from those described in this presentation due to a number of risks and uncertainties. For factors that could cause Marvell's results to vary from expectations, please see the risk factors identified in Marvell's annual report on Form 10-K for the fiscal year ended January 29, 2022. Let me now introduce Dan Christman, EVP of our Storage Products Group, who will kick off our CXL discussion. Dan has led our storage products for the past 5 years. He has been instrumental in pivoting our storage franchise from a majority consumer business to now a leader in data center storage. Under his leadership, we have entered new markets in storage and finding new business models such as do-it-yourself SSDs. With that, I'll turn the call over to Dan. Dan?

All right. Thank you, Ashish, and great to see everyone today. So I'm going to review our Storage Products Group today and how CXL fits in with our overall strategy. So as Ashish mentioned, we've spent the last 5 years optimizing our storage portfolio really for the data center. And today, this data center actually contributes more than 60% of our revenue in our storage business. Our storage business is made up of 3 business units. The first one being our fiber channel business, where we are the leaders in fiber channel controllers and HBAs. We also continue to invest in this market with our enterprise partners. Our second business unit is our HDD and of course, Marvell was founded as an HDD company. And we continue to be the market leader in HDD controllers. Again, here, we transitioned our business from a larger consumer business really into a data center focus business for HDD controllers, and we continue to see tremendous traction here as well. We also have a meaningful penetration now into the preamplifier market. We've continued to up our share and this was a new initiative from Marvell a few years ago at the behest of our customers, and this continues to ramp in revenue, it's doing very well. And then our third business unit is our flash business unit. And this is where we've really been extending our leadership in SSD for the data center SSD controllers. Last year, we announced our Bravera branded data center solutions, and we continue to win with the major NAND vendors with our PCIe Gen5 and Gen6 SSD controllers. We also have a flexible business model in our SSD business, which spans from merchant solutions to ASICs, all the way to full [indiscernible]. We also include full turnkey firmware, which helps our customers optimize their solutions. And that leads to multigenerational design wins. We also leverage our 5-nanometer platform in Marvell and collaborate across Marvell to integrate world-class compute IP. This led us to deliver cloud optimized storage accelerators, which continues now to expand our business even further. But today, we're here to talk about the CXL. And we'll talk about -- this is the next growth vector, not only for Marvell's story, but Marvell as a whole. This is an exciting technology. It's a new protocol really that opens up the connectivity of memory over the PCIe bus, which is particularly PCIe Gen5 and PCIe Gen6. This really represents an expanding market opportunity for Marvell across many of our businesses, not just storage. And at the company level, we've been investing in CXL for some time now, and we'll spend the next 20 or 30 minutes, really going over this opportunity and why we believe Marvell is the best positioned company to lead in CXL. Next slide, please. Now as we've discussed previously, Marvell is uniquely positioned to win in storage. But the strengths and capabilities that we have in storage translate over to the memory space. We believe we can also be the leaders in CXL or memory composable or composability via CXL. Now let's start with technology leadership. Marvell's been in storage for 25 years. We've incredible IP in storage, in controller technologies. And we also advanced 5 IPs, right? In this -- and CXL leverages the PCIe interconnect, having leadership in, for instance, PCIe Gen5, PCIe Gen6 is critical to be successful in CXL. We also have access, as I mentioned before to all of Marvell's IP to really provide cloud-optimized solutions, which is really the tenet of our strategy and our platform there. We also leverage Marvell's 5-nanometer silicon platform for our CXL solutions. That puts us 2 generations ahead of our competition and provides lower power and higher performance to our customers. We have flexible business models, which really allow us to engage with our customers, which is really kind of, again, for cloud-optimized silicon, we need to be able to work with our customers across, whether it be merchant, whether it be ASIC or fully customized solutions, and we also pioneered the DIY model in SSD, which is critical because that's actually created an opportunity for us to work closer with the NAND vendors and get early access to the NAND and technologies, but those same customers also make the DRAM, right, with the DRAM is what's important for CXL. And then lastly, we really created a world-class firmware team over the last 5 years. And this is really a data center firmware team. It allows our customers to optimize their solutions and provide the optimal solution for their infrastructure. This equally applies to our CXL products. Now as I mentioned, we've been working on CXL for some time now. Given Marvell's comprehensive data center IP portfolio, our cloud-optimized silicon road map and our position with the cloud data center customers, this has truly been a priority for Marvell. And now with the acquisition of Tanzanite, this only accelerates our efforts. Tanzanite is a leading provider of CXL switching fabrics. They've been focused on CXL expanders, CXL pooling and CXL accelerators. And this fits right into the overall Marvell strategy to lead in end-to-end CXL solutions for the data infrastructure. So I'm really excited about our talk today, and I'm really excited now to hand off this presentation with Thad Omura, who's our VP of Marketing for the flash business unit. Thad will talk to you about the opportunity, the incredible opportunity for CXL, and then also kind of why Marvell is uniquely positioned to lead in CXL. Thad's an accomplished 27-year industry veteran. He's been at the forefront of emerging technologies, taking them to mass market adoption, including RDMA networking technologies as well as solid-state storage. Now he's driving Marvell's efforts in [ evangelizing ] CXL across the entire industry, including the data center customers. So Thad, with that, why don't you take it away.

Awesome. Thanks, Dan. So I'd first like to set the stage and cover what memory challenges we constantly hear about our top cloud customers constantly share with us. You may or may not know, but DRAM is the largest component spend in the entire data center; that's more than NAND flash, it's more than CPUs. It's not surprising how memory right now is a major focus today in optimizing the usage and utilization. So when we look at how DRAM is connected into the infrastructure, you'll notice it's always directly attached to a processor like a CPU, a DPU or GPU, and that connects in through an integrated DRAM controller. Now this makes it very inefficient to share. And when another server may want access, it's extremely difficult to gain that through another CPU. As a result, you have a lot of expensive memory that simply goes unused or is underutilized. And cloud vendors are looking for easy ways to improve this. On the other hand, it also means if your workload has plenty of compute resources but only needs more memory, you are forced to add yet another CPU you don't really need. It's like these very power-hungry CPUs are an expensive tax just to add more memory to your infrastructure. This is a problem we continue to hear about. The next issue is that the core counts of these state-of-the-art processors are increasing at a much faster rate than memory bandwidth can serve. You have more and more powerful CPU cores that are starved for data and this degrades application performance and efficiency. What you see here is a chart that Meta presented back at the OCP conference last November. And what it shows is the relative CPU core count increase over time. That's the top line. As we know, this is on a fast rate of increase. The bottom line is the relative memory bandwidth for core. And you'll notice this is declining because DRAM performance just can't keep up. Now what is really concerning about this chart is the gap continues to increase over time, which means your very expensive CPUs are actually becoming more and more inefficient at processing memory-intensive data workloads. Clearly, cloud data centers are motivated to find a solution to reverse this trend. And another major challenge is implementing the so-called near-memory compute acceleration. To date, we have seen major benefit by adding acceleration, heterogeneous compute, domain-specific compute functions to network adapters like SmartNICs and DPUs, and we've seen acceleration move closer to storage for IO virtualization and security. But to date, we have seen no easy way, no standard way to add compute acceleration directly to the memory devices themselves. As a result, memory and compute-intensive applications that are looking to scale performance have suffered. Now CXL is poised to address all of these challenges, all of these issues, okay? So let's jump into what is CXL. CXL stands for Compute Express Link. And this is an industry standard protocol that runs over PCI Express, which is the most ubiquitous in-the-box data center in interconnect. What makes CXL so interesting is that it really focuses on the low latency transactions. So you can efficiently carry memory transactions from a processor to a device that is attached. It also builds on the ability to be cache-coherent. Now what this means is that data that's being processed by the host CPU, it's the same copy as the data sets being processed in the attached accelerator device, and through this protocol, you can enable simultaneous processing on the same data. This allows applications to run orders of magnitude faster. There are different types of CXL devices that the standard specifies you see on the chart here accelerators, you see accelerators with actual DRAM attached to them on the device. You see memory buffers where these are devices that find a way to add more DRAM into the infrastructure. I'm going to cover all of these applications and all of these different CXL types in this presentation. So let's quickly recall all of those cloud data center memory challenges that we talked about upfront. And what we're going to do now is shift gears and focus how CXL-based solutions address all of these challenges. We're going to cover CXL expanders, okay? They solve a couple of different issues. We're going to cover CXL cooling devices. Switch solutions as well as accelerators. Okay. Let's dive into the first application of CXL, which is commonly called a memory expander. Now traditionally, a DRAM DIMM module must plug in right next to the processor. There is limited space due to thermal and distance constraints. The CPU is generating a tremendous amount of heat and this limits your scalability to add more and more DRAM capacity. Also, you have to completely open the server to plug the DRAM in, which means it's not really serviceable once the server has been deployed. And because only the DRAM memory devices are on the DIMM, there's no telemetry information. Like let's say you wanted to understand how long the DRAM DIMM has been in service or how many errors have occurred over the last number of years that needed to be corrected, you can't easily keep track of this data. So what I'd like to now do is introduce you to a CXL expander. This is a module that puts DRAM memory behind a CXL expander controller. Because there's a controller right on the module, you get a number of benefits. You now have a module that you can actually scale performance and capacity because it can be pluggable. You can add it right into the front slot of the server. Keep in mind these are the same slots that are already being used today to plug in SSDs. That means it's serviceable, if there's a problem, right? You can pull it from the system without powering off the whole server and opening it up. You can add telemetry to the expander because a controller is directly attached. And now you're much more scalable because you can plug in as many modules as you need more memory. Now there's even more benefits such as improved thermals as these modules are sitting in a much cooler part of the server. There's a lot of high-performance fans that are cooling the front part here, and that helps increase performance and capacity because you can run the modules potentially at a faster speed. Another big value is you can mix and match DRAM technologies in the platform. So maybe you want to use today, DDR4 is a cheaper memory technology, but your server CPU only supports the latest DDR5 controller. Now you can plug those DDR4 modules directly into the front. We are also hearing you may actually want to recycle older DRAM modules from servers that have been retired for applications that simply need more capacity. So what's happening here is the DRAM technologies are no longer being tied to what the CPU DRAM controller supports. You can have configuration flexibility in how much DRAM do you want to add to the server, meaning you can decide much later in the deployment site since it's pluggable. So for the first time ever, you have an architecture that supports a pluggable DRAM module that are just as efficient to talk to as DRAM DIMMs with so many other benefits. So now you know what a CXL expander is, let's talk about how they actually improve performance. In a traditional service system, to get maximum DRAM performance to those CPU cores, a common configuration is to load up as many DIMM modules as CPU memory controllers you have in the device. So in this example, we're showing 8 modules are being plugged in for a total of, say, 128 gigabytes of data. But now you need to add more capacity, so you plug in 8 more DIMMs and what you have is a configuration that the industry knows is 2 DIMMs per channel. Now the problem is you pay for twice the amount of memory but you are getting the same memory performance and bandwidth as you did in your 1 DIMM per channel configuration because you're still connecting through the same number of DRAM controllers. Now this is a big challenge in getting more data bandwidth to an ever-increasing number of CPU cores. With CXL, we see the ability to use the initial 1 DIMM per channel configuration, right? You're still going to plug in, say, those 8 DIMMs into the 8 DRAM controllers on the CPU, but when you add in additional DRAM capacity, you're now going to do it with CXL memory expanders. What you're doing is you're utilizing the PCI Express ports of the CPU to open up more memory bandwidth as you add DRAM capacity. This means with the same capacity configuration as the previous one, you are getting much more data bandwidth to the cores, which significantly improves memory utilization, application performance, and this is a major breakthrough now in the CPU and server architecture, which will really allow application performance to scale efficiently using the latest state-of-the-art CPUs, as we all know, all increasing in CPU port count. Okay. Now we're going to go beyond expanders. We're going to look at new emerging use cases of how CXL will be used to further improve DRAM utilization in the cloud, not just the performance, okay? So let's first start with this concept of pooling memory over CXL. This solution, this architecture allows you to actually share the same DRAM across multiple completely different processors and different services. In the diagram, what you see is a processor can talk locally to its DRAM controller and soon, you'll be able to talk to a CXL pooling device that supports multiple host ports and you'll be able to do that without a real latency penalty. It means you can now easily rescue underutilized DRAM that is part of the CXL memory pool and dynamically reassign it to another CPU that really needs to take advantage of it. You also have an architecture where you can simply add DRAM memory to the infrastructure without having to add more expensive CPUs that would be redundant for purely memory-intensive applications. This is a tremendous value for hyperscalers looking to really optimize DRAM utilization. Now Tanzanite, which, as you know, is a company that we announced the acquisition of earlier last week, they are out in front by already demonstrating this CXL memory pooling capability with a tremendous amount of industry excitement. In addition to CXL pooling, you are also going to start to see the emergence of CXL switching solutions. Now with this architecture shift, you'll now start to see CPUs no longer integrate DRAM controllers on them at all. Instead, all DRAM memory will plug right into the CPUs -- right into the CXL switch through expanders and pooling devices. The CXL switch will be enabled to plug in all sorts of resources like IO devices, network adapt, storage. The resources will be able to scale completely independently, and now you can instantly mix and match resources together to rightsize their workload from the full pool of resources. This is what we call composability. And we're going to cover how CXL enables a fully composable data infrastructure in some follow-on slides. But before we do that, I wanted to touch on one other class of CXL devices because you saw that on the CXL spec slide that we call accelerators, CXL accelerators. We hear from our top cloud customers that in order for compute and memory intensive cloud applications to scale more efficiently, near-memory compute solutions will emerge in the form of these accelerators. So when you think of this concept, think about adding application-specific compute engines right into a CXL expander control. These engines can be CPUs, they can be dedicated processing hardware that operate on data right in the memory expander itself. This is, of course, completely coordinated with the host CPU. This significantly accelerates workloads like analytics, machine learning, complex search algorithms and the whole goal here, deliver faster time to insight for users. Accelerators are poised to dramatically improve the efficiency and TCO of the entire cloud infrastructure. Now continuing on this theme of acceleration, we're going to see more and more I/O devices leverage CXL. We're seeing DPUs and NICs on the network side, and we're also seeing SSD controllers on the storage side start to explore how do I take advantage of CXL through coherently accelerating protocol processing along -- in conjunction with the host CPU and what it needs to process? And we also see adding CXL to these devices will enable them to be fully composable resources in the CXL infrastructure. Okay. So that was a lot on the CXL products. And while you heard about all these like cool new capabilities and functions, what we want to do now is zoom out and talk about what Marvell's vision is of the data center infrastructure at a rack level. So today, individual servers typically contain a bunch of different component types, right? Compute, memory, storage, accelerators, I/O devices. They're all interconnected with Ethernet and they interconnect all the different servers within the rack. So in this architecture, there are servers that are especially designed to disaggregate storage. We are starting to see the emergence of this and accelerators that can also be disaggregated over Ethernet. And these resources can now be dynamically allocated to different workloads. But the emergence of CXL and cloud infrastructure means you will first -- you'll start to see the beginning of taking advantage of disaggregated memory and CXL accelerators. So these components still may reside in the server, but over time will start to migrate into their own appliance. And you'll start to see Ethernet provides really the connectivity between the servers and the rack, but there will be a low latency CXL fabric that can be utilized to enable more and more composability of memory in the rack. And then if we take this one step further. The rack architecture of the future will fully utilize CXL as a low-latency fabric. You'll have completely disaggregated resources that you can instantly compose to meet your workload needs at the click of a button. You may need more memory for some very large in-memory database or you need some massive amount of compute for an AI training workload. Think of all of the resources in the rack as LEGOs. You can basically quickly snap together the right capacity of any resource to support any specific workload. And when you're done, you just simply return the resources back to the pool and they're instantly ready to be assigned to the next workload. Again, you can scale any of these resources independently. They're not tied with each other. This full composability of the cloud data infrastructure is really the Holy Grail that operators are looking for. It makes the entire infrastructure as efficient as possible and as flexible as possible for all or any types of workforce. Okay. So now if we go back and take a look at those major challenges cloud data centers are facing in regards to memory, you can now see how CXL solves these problems. CXL is enabling disaggregated memory solutions or a fully composable data center. CXL is opening up the DRAM bandwidth to more and more powerful processing cores by utilizing PCI Express. This enables optimized application processing efficiency, and CXL is enabling near-memory compute solutions that will deliver the ultimate in application processing performance. We, big believers that CXL is disrupting cloud data center architectures, we see this happening today. So now you are familiar with the different types of CXL products that the industry is looking at. It's important to understand the technology road map and how we see the cadence of deployment. So as I mentioned, we're really just at the beginning stages of CXL solutions going to market. Server platforms that support CXL are just starting to emerge, and CXL solutions will follow that need to prove the value proposition and also be qualified in these systems. So right now, what you see emerging in the market, supports the earliest version of the spec we refer to as CXL 1.1. This being said, there is a tremendous amount of work happening today by cloud customers to architect their infrastructure and drive solution requirements into the industry. Marvell is heavily engaged with these -- in these discussions. We're working with the major cloud vendors as we know it's super important to gain deep understanding right now of the requirements, so that you're in the best position for volume deployment. Now from a spec perspective, CXL 2.0 is already released and this version of the spec, to find additional capabilities of CXL for wider application deployment. As platforms and product support this version come to market, you'll start to see different applications go beyond expanders and start to move into pooling and switching and accelerator solutions. And we may even start to see the emergence of I/O devices that take advantage of CXL. This is where we'll really see volume deployments and they'll be aligned with the next generation of PCIe Express Gen5 server platforms that are deployed. And then following 2.0, you'll see solutions based upon CXL 3.0. This further extends the protocol. Work on the 3.0 spec is well underway. These solutions will be deployed with PCIe Express Gen6 server platforms, and this is where you really start to see the future vision of a fully composable cloud data center come to market. Okay. Now hopefully, you've got really good exposure to all the different types of CXL infrastructure solutions that will impact at the cloud. And now what we'd really like to share with you is how is Marvell taking a look at this? How are we taking a comprehensive end-to-end solution approach to CXL? So we just reviewed a number of different types of solutions, including the expanders, pooling devices, switches, accelerators. These products have a tremendous amount of focus today. And our cloud-optimized approach enables us to engage with customers on solutions that require maybe some special hardware and/or firmware features for their particular infrastructure. What we also see is CXL leveraging the rest of our portfolio. This includes custom compute solutions as well as our networking DPUs and SmartNICs and our industry-leading data center SSD controllers. We discussed how leveraging CXL technology into these products will enable them to be completely composable, and optimize protocol processing with other compute devices in the infrastructure. Now what may not be completely obvious is based upon what you've seen today, but our copper SerDes and industry-leading electro-optic technologies are absolutely critical to realize all of the value that I presented to you on CXL. If you think about in-the-box high-speed signaling as well as connecting CXL-enabled boxes in the rack over maybe copper cabling, PCIe Gen5 and Gen6 signaling rates demand the absolute best state-of-the-art bi-layered technology to ensure low latency, high signal quality, and this is an area where we have proven leadership. At the same time, you may want to transmit these extremely high throughput CXL memory transactions over longer distances, maybe rack-to-rack, and we're in a great position to leverage our electro-optics leadership technology to ensure CXL has the best possible distance, latency, cost and performance over fiber connectivity. There's a lot of trade-offs to consider as cloud vendors architect their CXL-based solutions for these data centers of the future, and we're right there with them to optimize those enabling solutions. So as you take a look at the full picture of all of the technologies that are really required to deploy these solutions, you can see we're in a unique position because we have market leadership, proven execution in all of these critical data center disciplines. Because CXL is focused on optimizing memory transactions at the nanosecond latency level, it's really a critical factor in determining how effective your CXL solutions are only by measuring how good are they operating end-to-end. Now we are focused on not just optimizing at the digital level, but there are 5 technologies at the analog level. And having this all under the same roof means we can provide the most value to cloud customers who need a trusted partner to work as they architect and eventually deploy these CXL solutions. The acquisition of Tanzanite helped bolster and accelerate our road map, and it's a strong testament to our commitment to the CXL opportunity. Customers and market interest, I got to tell you, it's just increasing. Day by day, we see more and more interest in this technology. And given the impact that CXL will have to our entire portfolio, we absolutely believe this represents a multibillion dollar opportunity. So thanks so much for spending part of your day with us to get an update on CXL technology. I think you will see that CXL is disrupting cloud data center architectures today. We are thrilled to be working with the leading cloud vendors to drive solutions around this technology forward. Marvell is uniquely positioned to enable end-to-end solutions in the data center and we're powering the transformation in the cloud to benefit the most from the CXL technology. CXL is a natural extension of our cloud-optimized portfolio to address now the memory challenges and enable a fully composable data center. And we at Marvell see the CXL market, we're just at the beginning of a multibillion-dollar opportunity, including new CXL infrastructure solutions as well as adding value to our entire existing portfolio. So thanks again. I think we're going to now open it up to Q&A. Maybe, Harlan?

Yes. So thanks for that great overview, Dan and Thad. So we've got about 15 minutes for questions and have compiled a few. So I guess to start off, you talked about sort of the road maps for CXL deployment, starting off with CXL 1.1 protocol moving to CXL 2.0 protocol. My understanding is that these are tied to the microprocessor rollouts by AMD and Intel. So my understanding is that CXL 1.1 will be rolled out this year with Sapphire Rapids from Intel, and I think Genoa from AMD and CXL 2.0 will start to roll out in 2024 with Intel's Granite Rapids and whatever product that AMD has at that time. Is that kind of the right time frame, like CXL 1.1 starting this year and next year; 2.0 in 2024?

Yes, Harlan. That's about the right time frame that we do see. Certainly, there's a lot being driven by the leading server vendors from a spec perspective, and then you'll see even in between though, people have some additional features on top of 1.1 even in the time period in between. But in general, your rollout is aligned with the way we're thinking about it.

That's great. And so you guys are building on a complete sort of, as you call it, end-to-end portfolio of stand-alone products, but you're also going to be integrating CXL and some of your core products, right, like core compute, DPU and some of your storage products. I guess so whether it's the stand-alone expanders, pooling, switching devices or CXL integrated into your core devices, when are we going to start to see Marvell start to roll out products in support of CXL?

Yes. We're going to be aligned with the time frames that you had mentioned previously. Today, we really wanted to focus upon sharing all of the key technologies that are going to be coming to market, and we're going to have some really exciting announcements as we come forward and prepare our product portfolio for announcement in the future.

Yes. Harlan, maybe I'll just add a little bit here, which is, as Dan and Thad mentioned, we've been working on this internally. And I think we felt this is the right time frame, right? So as you discussed, 1.1 is -- which is really the architecture phase, and that's what's in play today. So I think you're thinking of the right way, which is there's a lot of activity going on below the covers at this point in time. What we do today, right, over the next few quarters and next year or 2, really sets up what happens in this market for the next month or number of years. So this is the right time frame, why we came out now. And why we're discussing it now is because while there's been a lot of activity kind of happening from a market perspective in terms of getting the spec organized, things are now accelerated to where the decisions are being made, right? And that was one of the big reasons why we also accelerated, bought Tanzanite and get things, we just need more resources by time, right, given the level of activity. So you should expect product announcement, product rollout, stride, design wins, things like that will start to come. But we wanted to set the stage, right? This is a new technology, and that was really the purpose of today was just to get people familiar with what is this all about.

Perfect. So talk about the competitive environment because we've been hearing a lot more about the discussions around adoption of CXL. So for example, some of the big memory suppliers have been talking about potential developments on CXL, especially the DRAM suppliers. You've also got some of the IP suppliers out there that are focusing on this segment, and then we also have a number of different start-ups focused in this area. But from my perspective, like many of these companies are looking at it through somewhat of a narrow lens; I mean the memory and the IP guys are looking at it purely through the storage and memory angle, for example, but obviously, as you pointed out, CXL is much broader than that, right? CXL is about pooling, it's about interconnecting compute and acceleration ASICs and not just memory, but storage. And so these are all sort of the key silicon engines and building blocks that Marvell is focused on. So what's your sense on how much of a competitive advantage you guys have? Because not only will you have stand-alone products, but you'll be integrating CXL across your compute storage and networking products.

Yes. So exactly what you said is that as we are deeply engaged on the architecture and how these things will be put together, a lot of our competitors or a lot of other folks looking at these technologies may be looking at one particular part of the infrastructure to optimize where we're taking a much more holistic view and making sure that if you have an expander that's directly attach, how does it work if all of a sudden, it's completely disaggregated and now the system? How are you going to connect to it most efficiently? And the fact that we -- under the same group, I can discuss with my partners, BUs that focus and have industry-leading technology on the connectivity side and then I'm engaged with the customers on the digital expander pooling devices. It really brings together a complete solution, which is really what customers are looking for, solve for me end-to-end how this is going to work. So that's where we do think Marvell brings a tremendous amount of value up and beyond the majority of our competitors also going after the space.

And Harlan, just to add, I think part of it is, if you think about the amount of touch points we have within cloud customers and how that's expanded like massively over the last couple of years, right? Because remember, there's a huge systems architect sitting on the other side of the wall, right? And the fact that we have multiple touch points into them and we can talk their language, right, for each of those different domains, I think it's a big differentiator for us, right? That's why we see ourselves as being a very, very key part of this joint architecture over time with our key customers.

Great. And then help us compare and contrast CXL protocol and the connectivity and the interface architecture because at a high level, it kind of sounds familiar to what others are doing in the compute space with more maybe proprietary-based fabrics and they talk about like memory cache coherency and so on. So for example, AMD's Infinity Fabric, NVIDIA's NVLink and NVSwitch architecture and there's a number of other types of fabric architectures out there. Can you compare and contrast CXL to some of these proprietary-based solutions?

Sure, absolutely. Yes, there have been a number of other previous attempts to pull together a solution to really optimize memory transactions over fabric. There are other ones like C6 and OpenCAPI and even GenZ, which is now part of -- which is a part of CXL. The real difference is that CXL made a key decision and has been driven to run over PCI Express interconnect, which is, as we said, a ubiquitous kind of in-the-box type of interconnect. So you have a whole infrastructure, if you will, of solutions and testing capability and what have you to leverage on PCI Express for the physical side. That's one of the big drivers and a reason why, so has got so much momentum. And then on the -- just kind of on the industry support side, you're starting to see the major server platform providers announce and drive support into the market. You're seeing now the memory and the component type. So there's just a tremendous now industry momentum behind specifically CXL. And one of the things that has happened is the PCI Express protocol has now gotten fast enough from PCIe Gen5, moving to Gen6, where it makes sense now to put memory transactions over that interface. So all of that has kind of now happened and why there's such momentum behind CXL.

Wasn't apparent to me until your teaching today was that Marvell's networking connectivity leadership actually plays somewhat of a big role here, right, as you disaggregate compute and memory and storage, you can have, as you point out, pools of memory that are physically separated from the server CPUs and connected via these high-speed networking connectivity links, which means that you will have to develop CXL-based sort of networking connectivity silicon that keeps the latency to a minimum. So is that the plan for the team, is that you'll have a -- obviously, you have your leadership in optical and copper connectivity that supports Ethernet but then you'll be rolling out an entire new product line of networking connectivity that's based on the CXL protocol? Is that how the strategy works?

Yes. Again -- yes, go ahead, Ashish.

Yes. I think -- Harlan, you're thinking on the right way, which is, yes, it's going to -- this is going to impact pretty much all of the product line, right? So for example, as you know today, right, within racks, for example, for enabling gearbox functionality, which is connecting lower speeds via higher speed connection, right, you start to see the same things product type, right, the industry work typically, things like [ gate handles ] and gearboxes, right? Those products will basically start to get embedded, right, with CXL technology. That's exactly what we start to see.

Perfect. Well, we're just about out of time. This was a great overview of the CXL standard and connectivity technology and Marvell's focus on capturing what I think is first-mover advantages at the silicon, software and firmware level. So I want to thank Dan. I want to thank Thad and Ashish for this very informative session, and thanks to everyone out there for participating. Hope everyone has a great week.

Thanks, Harlan, for hosting us.

Thanks, Harlan.

Thank you.