One Stop Systems, Inc. (OSS) Earnings Call Transcript & Summary

Hi, everyone. Thanks for joining us today. The webinar will begin in just a moment.

Thank you, Katie. Today, we're here to talk about turning a large data set intellectual property into AI Gold. So today, it's Jim Ison here from One Stop Systems, Chief Sales and Marketing Officer. We'll hear from Matt Hallberg from KIOXIA in a little bit. And moderating your questions will be Tim Miller of OSS. So the agenda today is to talk -- start off talking about the AI transportable marketplace and the challenges of bringing edge data into solutions that can be managed through AI workflows. After that, we'll talk about -- Matt will talk about KIOXIA drives and how the CM6 and CD6 drives are ideal for local and remote AI workloads and talk about the encryption capabilities. Then we'll talk about successful edge solutions that OSS has deployed with these drives and systems. [Operator Instructions] So if you look at any research company out there that's talking about edge computing, they're all talking about explosive growth. There's a few here. It's specifically Zion Market Research, interesting on the data side is, in 2018, only 10% of the data that's generated at the edge was actually then processed at the edge. But as they see it by 2025, they expect 75% of that data that is captured at the edge from sensors to be processed by high-performance computing at the edge. So we'll talk more about that when we get into the AI transportable marketplace, which we'll start here. One Stop Systems has a product line called AI on the Fly and how does that get you to this AI Gold. Really what AI on the Fly is to get into actionable intelligence when things must take place now and don't have time to give back to the data center. So there's a group of vast data sensors out there that are capturing data. So what we need to do is, first, get that recorded, using high-speed data acquisition, all that IO. After the acquisition, we need to store that data and make it available to various compute resources. So it needs to be low latency, high throughput and NVMe is the preferred path to make that happen. After it's stored and made available, then it needs to be computed. So using multi GPU, FPGA, compute engines and algorithms for AI frameworks, all those inferencing, training and retraining has to be done on that data. And really what that's to do is to create actionable intelligence. That's the gold. What is those actions? Well, if you're in an autonomous vehicle, it's how to react to something new that's happening in your environment or when to turn. When you're talking about something like an oil rig, it's where to drill to get to that black gold, so to get to the -- which you're after and having artificial intelligence to do that. And if you're in an autonomous mining equipment, it's where to mine to get the actual gold. So we'll take a look at the edge computing marketplace. You've seen those research analysts talk about it, but it's described and segmented in many ways. Usually, the cloud is at the top. That's a hyperconvert or a hyperscale data center that is somewhere in the world generally remote from where the data is actually being gathered. There's some edge data center, edge nodes, called FOG nodes sometimes, depicted that way, and edge devices like IoT devices out there. So since there's so many ways to describe it, we'll try to put it in perspective for how the AI transportables is a piece of that market. So there's still the cloud and central data center. That usually has Tier 1 OEM hyperscalers by those Dell, HP, IBM, those type of products and having an air-conditioned building. Those companies also have what they call edge data centers. Well, still an air-conditioned building, just happens to be maybe closer to where the data is generated out in industrial areas and things like that, trying to get closer to the data. Why are they doing that? Because there's tens of billions of high IoT devices out there, anything from game consoles to your refrigerator, to wearable devices that need using 5G usually access those data centers, one way or another, to share information. What we're talking about here is really the AI transportable market. So those are things where you need the power of the data center not in an air-conditioned building, but out on the edge, in a medical device, in a truck, in an aircraft, something like that. So that's really the focus of what AI transportables is. And those are definitely edge applications. So some of the industrial applications we're talking about that require that high performance, anything from autonomous vehicles, and most people think about cars, but it could be trucks, buses, trains, anything that moves. Aviation, commercial aircrafts generate so much data that is not really used because the satellite link is relatively slow compared to what we're used to in terrestrial networks, and the black box just doesn't store that much data. It's not what it's paid for. The oil and gas industry uses high-performance computing to find out where to drill for oil, so they don't waste money drilling in places where there might not be oil. And even in media and entertainment, doesn't seem like a high-performance application, but follow a set from place to place or a follow U2 or Katy Perry or somebody from location to location and the high-performance computing that make all those great visual effects needs to be organized in order to keep that -- keep at the edge, keep moving around. Other industrial applications from video analytics to medical devices and even lighting for gold and creating food, all have edge applications that could all be done with autonomous surgery machines or autonomous vehicles in those applications. So as far as products go that do that, we talked about the ingests that getting all the data sensors into and recording those. All-flash arrays using NVMe are the ideal platform to do that with. We have the high-speed sensor inputs, need multiple power choices at the edge, can just be a standard wall power, might need to be a DC power from a battery or generator or something like that. They need to be ruggedized for the environment and include the recording software, like what we're showing here is the semi-rugged platform, the DRC product, that's an OSS' data recording platform where we can actually record high-speed data directly to the discs using a bunch of sensor inputs, including FPGA-type IO devices. Fully rugged platforms like the FSAn are another storage that has -- you see the removable canister concept at the top, where we have 200 to 400 terabytes of data that can be stored and recorded on those type of ruggedized platforms. After they're recorded, you can use that same flash storage array, the FSAn, to do your data storage and serving to several edge-type computing devices. So all-flash arrays still make up the majority of what you need at the edge to stay rugged. And at this time, you add SAN, NAS or even direct attach-type -- Internet software and direct-attached hardware to get the highest speed available. NVMe over fabric makes that super simple to be able to share storage data between devices. And usually, you have 1 or 2 of these out in the field. So they need to be high availability in addition to rugged and also be scalable to do massive amount of data that's being collected and needed to do -- train your AI workloads, your AI frameworks. So the Ion accelerator all-flash array, the FSAn [indiscernible] has that software available for SAN or NAS in a rugged-type environment that can be flown in an aircraft or put on a truck. The SB2000 is another rugged platform that has features like being able to pull the -- 8 drives at a time in that canister or making the drives come out individually like most JBOFs are. Then we have the edge server. So once you've got the data, you've got it ready to be processed, you need the servers to do the processing. So the compute we have for the edge, rugged and servers, here you can use the servers individually to scale up and do a large-scale inferencing. Composable infrastructure-type products, when you're in more of a rack scale-type of solution. Using AMD, ARM or even Intel solutions for -- on these servers can all go into these rugged enclosures along with the software, the AI frameworks and management software to make it all easy to manage even at the rack level, short depth, and when we say no compromise solutions, we're talking about having the highest performance GPUs, the highest performance CPUs available for the application, but putting it out, like I said, the power of the data center out at the edge. The EOS server here is our Gen 4 server platform that has Intel AMD and future can be ARM processors with a scale out bios that we'll talk about in the next slide. The short-depth server has those same removable canisters as we saw in the SB2000, so that can actually be moved from one product to another, but also has more hyperconverged characteristics where you can put 16 KIOXIA drives in there and GPUs -- 4 GPUs, FPGAs and different power solutions to be edge compatible. If you need more scale and inference, well, we have expansion systems. So that's where you want to scale out many inferencing platforms from a single server. That also comes with the same software management. And what we say tunable performance is we could put the power of the data center at the edge. But if you have a temperature requirement or something like that, we have ways that we can actually power cap the GPUs as you can hear from Matt when he talks about the drives have the same kind of capabilities to make a perfect fit solution. By this, we have a 4 way -- the EB4400 is a 4-way A100 GPU solution for Gen 4, has up to 28 inferencing engines in that box alone. We have a full rack mount version that has 8 of those. So you get 56 inferencing engines in that solution. If you like something like the NVIDIA T4, which is smaller, maybe lower power, we have scale out solutions that can have up to 32 of those T4s in a single box that can do a lot of, say, video analytics-type applications because of the encoder decoders on those T4s. So you see we can scale out as well as scale up with our GPU compute, all at the edge. And when you need high performance, the GPU-accelerated system, like the GAS-R, give you that capability in a very rugged form factor, can go on an aircraft, in a truck, has 8 of the latest SXM Gen 4 GPUs in it for large-scale expansion, giving you massive performance per watt, all ruggedized to the environment and high performance. The GPUltima is a rack-scale solution, where you can rack up multiple of any of the rugged servers you've seen here and maybe put in a mobile data center, which we'll talk about in a little bit. So at this point, I'd like to turn it over to Matt to talk about the KIOXIA drives.

All right. Thank you. So just a couple of legal things to get out of the way. This webinar is for information purposes only. The information presented by KIOXIA and myself in this webinar is for informational purposes only and is not an endorsement of any company nor an endorsement of any use of our products. For unintended usage restrictions, we restrict to sale of KIOXIA the products and equipment or systems that require extraordinary high levels of quality and/or reliability and/or a malfunction or failure or -- of which may cause loss of human life, bodily injury, serious property damage and/or serious public impact. Because of proposed applications and use of products for which you -- which requested samples may be an unintended use, KIOXIA will request additional information about your application in use to consider whether or not to sell the products to you for the application of question under its guidelines under the unintended use policy. Therefore, please be advised that as a result of its unintended use consideration, KIOXIA may decline to sale the products or support your production stage for said et application in use or to sell you the products or to support your production stage only under certain conditions. Lastly, for Export Control Compliance, product software and/or technology received directly or indirectly from KIOXIA may be subject to certain U.S., Japanese, and/or countries, areas, export control laws and KIOXIA's export control appliance program. Customers are responsible for adherence to such restrictions, including, but not limited to, the following: none of the products be -- knowingly be sold, transferred or otherwise made available directly or indirectly to any 4 -- to or for any military end users and military end uses; none of the products, software, technology derived for the products will knowingly be sold, transferred or otherwise made available directly or indirectly; as an element of the design, development, use or manufacturing of any application specifically developed, configured or adapted for military purposes or any weapon or military extensive purposes; none of the products or product, software, technology derived from the products will knowingly be sold, transferred directly or directly; as an element for the design, development, use or manufacturing of nuclear, chemical or biological weapons or missile technology products or to entities identified in the U.S. Denied Persons lists, Entity lists, Specially Designated Nationals list or other lists, which may be published by the U.S. government for the purpose of prohibiting that entity from receiving U.S. goods or technology or to any individual or entity that violates any of the above prohibitions. Products received from KIOXIA or derived from KIOXIA products will not be exported or reexported from your home country unless authorized by law or by specific written authorization. Okay. Sorry. So my name is Matt Hallberg. I'm the Product Manager here at KIOXIA. I have a couple of slides that talk about why local storage at the compute node, remote storage, et cetera, matters, and how we get to that AI Gold. So picture here is an OSS FSA 4000 and an OSS HGX-2 Rugged system. And for the compute nodes, there are a lot of different workloads that are compute-intensive. I chose to just focus on a couple. So for instance, the training phase of machine learning is the most resource-intensive set of operations. For data sets, these are growing in size at a very fast pace. For those of you who may have gotten an MRI lately, each MRI image can reach multiple terabytes, and a training set of these MRIs can be composed of thousands of images. So we're talking about thousands of terabytes for a training set. And whether you're running on the RAM or on local storage, let's say, the OSS HGX-2 Rugged, the local storage needs to be able to handle reads and rights in blazing fashion with little impact to overall latency. Moving the data in, moving the data out, check pointing, all these things need to be completed quickly to minimize the idleness of the GPUs. As at the end of the day, the most exposed -- expensive component of your system is going to be the number of GPUs you have in it. And if the GPUs are idle, then you're not making use of them effectively and you're sort of losing out on your investment. For PCIe Gen 4, SSDs have very noticeable benefits with applications or workloads like file copying and other IO tasks, especially versus PCIe Gen 3 SSDS. So for our drives on sequential workloads, we see 6,900 megabytes per second on reads, 4,200 megabytes per second on rights. If you're doing random workloads, our drives are all around 1 million-plus IOPS on random reads and then 70,000 plus on random rights. That 70,000 plus is on our 30 terabyte drive. If you were to look at our 15 terabyte or our 12.8 terabyte drives, the random rights are ranging from like 150 to 190 all the way upwards of 300,000 random right IOPS. PCIe Gen 4 SSDs are also able to take advantage of 3D NANDS, higher densities, allowing for up to a 30 terabyte drive and a 2.5 inch SSD. We actually offer 30 terabyte drive and it's enabled through companies like OSS. And then lastly, cool things coming are already on their way to the compute nodes is if you are using GPUs from NVIDIA, they have this cool new technology called GPU Direct, and that's via Magnum IO, and that allows the GPUs to talk directly to local storage, bypassing the CPU and DRAM. So what that means to you is, in general, you have to have a pretty fast processor, pretty fast DRAM to have really fast interconnects between the 2 and to have banks and banks and banks of DRAM. And all DMA transfers to and from the GPU to the SSDs have to go through the CPU and DRAM. The CPU and DRAM are more or less staging, where they're taking the data in and then they're pushing the data out to the SSDs. But with GPU Direct, the GPUs can actually talk directly to the drives and do DMA transfers directly between them. So there's very little-to-no overhead on your CPU and DRAM, and that CPU and DRAM can be used for other tasks within the system. Okay. So this is a picture and a high-level overview of our CM6 product. This is our PCIe Gen 4 drive, NVMe 1.4 compliance, 2.5 inch, 15 millimeter Z-height. It's our own controller, firmware, flash, et cetera. It's available as single port and dual port. And a cool feature for us is this is our sixth generation of offering die failure recovery and double parity protection. That means is, is you wouldn't expect to see die failure at the beginning of the life of the drive, but as you get closer to the end of life and the drive has been worn quite a bit, if you have a die failure inside of the drive, the drive will be able to recover from the die failure and without breaking. So you'll see a performance -- a small performance hit because that's going to eat into your available over-provisioning space. And you'll get a smart warning, smart being one of the things in NVMe Sata and SaaS has it's own thing. But you'll see smart warning that a error occurred and that you should probably get ready to retire your drive. And because the drive is still fully operational, you'll be able to pull all of the drive -- all the data off the drive and migrate to new drives. We feature power loss protection and data protection. We have a bunch of different security options, non-SED, SIE, SED and FIPS 140-2. FIPS 140-2 is in process, which should be done in July. As Jim had also mentioned, there are 6 power mode settings. So you're able to tell the drive to operate under lower power in order to conserve electricity and reduce the thermal imprint. So if you're in a rugged environment that is high heat, you're able to tell the drive to operate at, let's say, like 14 watts to reduce the heat coming from the SSDs. And the operating temp for these drives is from 0 to 70C, and we have a feature called thermal throttling, where if we detect that the drive is getting too hot, we're going to slow the drive down from a performance outlook, and if it reaches a critical temperature, we're going to basically have the drive something like a [neck], where they would say, sorry, we're busy right now until the drive is properly cooled down and it will resume operations. So what that's doing for you, the end user, is you're making sure that the drives don't prematurely wear themselves out or die in the field because we're going to make sure that we never reach that critical state. On the performance side, you'll notice, as I mentioned previously, 6,900 megabytes per second across the board, except for 30 terabytes, sequential rights, depending on capacity. The bigger the capacity the better. 4.2 terabytes per second. On random reads, as I mentioned, over 1 million IOPS. I've been in the industry long enough to remember when storage box vendors were mentioning that their boxes were capable of million IOPS. And now you have drives themselves that are capable of 1 million-plus IOPS. On the remote side, remote storage, so your storage node or your data lake, like an FSA4000. Networking speeds, technologies and topologies have greatly advanced over the past few years. We've seen the advent of 200-gig mix. We've seen the proliferation of Rocky V2 or RDMA over Ethernets. NVMe over fabric deployments have started to catch on. And people starting to beef up basically what they're doing on the network side. That's going to be super helpful for when you have connections between your storage nodes and your compute nodes. As I mentioned previously, the data sets are composed of thousands of files, all of which need to be sent and received to and from the compute node to minimize downtime. So storage that you have, these NICs, these really fast NICs should be optimized for sequential performance to send data for processing and also to receive the processed data. And the faster the offload can occur from the compute node, the faster the local storage can send the data to the GPUS. NVMe over fabric deployments here show their strength by improving performance and reducing latency. And by using NVMe SSDs in the same environment for staging warm data versus cold data, that's going to optimize the spend that you have in remote storage. And lastly, GPU Direct technology can also take advantage of the remote NVMe SSDs. One of the unique features of GPU Direct is that it's also able to talk directly to the NICs. So it can talk to a NIC, the NIC would talk to the remote storage, and from there, they'd be able to access the drives, again, bypassing the CPUs. So this is our CD6, our data center or value NVMe SSD, also PCIe Gen 4, same controller firmware, NAND, single-port design, so useful for applications that require high availability. And it has the same die failure recovery and double parity protection, same number of power mode settings and same security settings. The performance here is slightly dialed down compared to the CM6, our CM6 being our flagship performance products. And then the rest of the performance numbers are in line with the PCIe Gen 4 SSD. Okay. Just a little bit about software-based encryption versus hardware-based encryption. So one of the knocks on encryption in totaling in securing the data you have at rest is the performance impact. At least that is a prevalent industry opinion that when you have security enabled, it becomes kind of a resource hog. And with the advent of newer processors, PCIe Gen 4 and whatnot, it doesn't have to affect, at least on the software side, as much as it used to. So in the diagram I have here, all the data that's coming into the system has to go through the CPU and DRAM, has to go to the software layer and then back to the CPU and DRAM. So what that means is the application layers using those system resources, the CPU and DRAM expressly, in order to encrypt and decrypt the data, which increases the system overhead and causes performance and resource constraints. Whereas if you have encryption and decryption directly located on the drives, it's built into the data path on the drive, it's encrypted and decrypted with no impact at the SSD and system levels. The other thing that I wanted to talk about since we are talking about security is, what is the difference between an ISE -- or SIE, Instant Secure Erase Drive, and a SED, a Self-Encrypting Drive. So the data path for an SIE drive, it passes through the crypto modules and data is written to -- the data that's written on the SSD is encrypted. When the data is read, it goes through a media encryption key and the 256 crypto module and is given to the requester as unencrypted plain text. So while the data at rest is indeed encrypted, if anyone were to read or right to the drives, they'd be able to read from the drive, all the data coming from the drive would be plain text, which isn't exactly securing your data. Whereas with an SED and with the FIPS drive, there is an authentication key and authentication credential process. That has -- that process has to be exchanged prior to allowing access to the data on the drive. So if the drive is not unlocked, if the credentials don't go through or if the credentials never are exchanged, you can't access the data on the drive. Okay. And then for FIPS, I'd like to just do a quick overview of what the FIPS process looks like and why it takes so long and why it's so difficult to find FIPS drives, and I wanted to just kind of put it out there that there is a difference between FIPS-ready and FIPS-certified. So at the vendor level, vendor like KIOXIA, right, we design and produce cryptographic modules and comply with the requirements specified by NIST. We submit these modules to a third-party, accredited cryptographic and security testing lab. And then any changes to the crypto modules after the validation means that we have to do a new submission. So this third-party lab, the CST lab, is accredited by NIST. They will independently test the cryptographic modules to the appropriate FIPS 140-2 and FIPS 140-3 levels. They'll conduct the algorithmic testing, review the documentation and source code and perform operational and physical testing of our modules. They'll then submit a written report to the validation authorities at NIST that the modules have met the requirements outlined in FIPS 140-2 and FIPS 140-3 documentation. And those test results are documented in the submission package by the CST. If there's any issues, the CST lab will work directly with the CMVP or with NIST. And then once the submission is received by NIST, they'll post it to their website. They have a CMVP module that you can go to at any time, and they'll post it as review pending. So there are 3 different stages of review certification with NIST for FIPS certification. There's in-review, coordination and certification. In-review is essentially, they revalidate the test results for each cryptographic module. Coordination is they're basically coordinating with the CST labs to verify that the results are accurate and complete. And if any changes need to be made, the CST will work directly with NIST to make those changes. Then lastly is certification. So if the test results and modules are determined to be compliant with FIPS 140-2 and 140-3, then the module is validated. Validation certificate is issued and the online validation list is validated. So this is a very resource- and time-intensive process. Resource for engineering, resources for money spent, time spent for the CST labs and the length it takes to go through FIPS certification. As an example, we submitted our FIPS 140-2 certification for CM6 and CD6 in July, and we're not going to have it done until July. So it's taken about a year. Previously, the process was around 6 months, but there were some government shutdowns and this whole COVID-19 thing that has severely impacted the resources and availability at the NIST accreditation facility. Not only that, there's been a new standard, FIPS 140-3, that's come about and everyone's submitting 140-2 now because the cutoff is in September. So that's one of the big differences between 140-2-ready and 140-2-certified is if you have -- if you're working with a government entity or a company that requires that the data is at rest and that it's encrypted, the only way for a government entity to be certified to be absolutely sure that this data is actually encrypted is via FIPS certification. FIPS-ready just means it could go through, but they're either -- they're not doing the resources or not spending the time to get it through or it's just in process, okay? With that, I'll turn it back over to Jim.

Thanks, Matt. Great detail on all that. What we want to talk about now is the successful edge solutions that we've deployed using the FIPS 140 modules and some of the OSS products and the KIOXIA drives. So one of the applications is a radio frequency applications that uses AI in a mobile data center-type environment. In this case, the customer puts out radio frequency antennas in the field, has a data center that's mobile, that collects the data from those using hyperconverged Gen 4 edge servers with the KIOXIA drives, records that RF data and processes it and the recording is happening at 40 to 50 gigabytes per second, even though we're doing the encryption, and that's at the full throughput. So that allows for fastest detection in the field of these RF frequencies. On the autonomous vehicles side, we have a path to level 5 driving. In this one, it's large-format trucks, not just cars. And with that, they need a high-performance inferencing platform that is separated. So in this case, the EOS server has the CM6 drives in the storage, but then, because of where it has to be in the truck, the GPUs, in this case, using the EB4400 sits under the passenger seat of the truck and needs to have that large amount of inferencing engines that are in there. So it's storing video and the sensor data, doing the real-time inferencing and has fast reaction to real-time changing conditions that happen on the road, whether you have to stop, turn, slowdown. And this is currently delivering a product in -- on the Texas roadways. And then some smaller applications that you would very high-speed needing to be done is in this 12K video recording. We talk about media and entertainment being an edge high-performance application. When you're recording 12K video at raw speed, you can see that amount of data that's generated just from shooting of TV shows or movies is a tremendous amount. So in this case, we take M.2 drives, put them in an OSS ruggedized canister that can then go into the cameras or be interfaced with the cameras to record that direct raw video directly to the drives. After that, we take the drives out of those canisters and using good old-fashioned sneakernet move those drives to a high-performance post-production system. With the amount of data generated actually moving it by FedEx or by hand is faster than trying to send it over even at 100 gigabit pipes. So high performance, fast dailies to post-production to get that movie and edited and on screen as soon as possible. So with all that, we want to thank you for today and see if we have any questions and answers, and turn it over to Tim, who will let us know what we have.

Yes. Thanks, Jim. We do have a couple of questions for both panelists. Let me start this one for One Stop Systems. Jim, do you expect your rugged GPU compute systems to end up in the self-driving cars like a Tesla, for example?

Yes, that's a good question. I mean we're talking about autonomous vehicles, but the word vehicle is fairly widespread. So Teslas, generally, have to do say -- I know it sounds like many things, but you're just inferencing a model that's already there. And then I believe Tesla has their 5G network to update all the cars with the information coming around. What we're talking about here is typically on a larger platform. So when you see these large scale out solutions, they're doing something more. I know driving is tough, but just driving, when you have to drive and deliver and you're doing in the truck application, you're doing more than just driving. And you've got to get that last mile to the loading dock and things like that. So larger -- you'd see One Stop Systems platform probably being used on larger-format vehicles rather than a Tesla.

Got it. Okay. This next question, probably applicable to both companies, but maybe, Matt, you could take it first. How has the adoption of the PCIe Gen 4 gone? And we're already hearing about Gen 5 on the horizon. Why does it make sense to invest in Gen 4 at this point?

Well, there have been a lot of delays on certain companies for getting Gen 4 systems out to market, semiconductor companies. So I wouldn't wait around on PCIe Gen 5. The movement from Gen 3 to Gen 4, for us, we were one of the first in the market with SSPs, and we've had really good success with -- working with companies like OSS and guys like AMD and NVIDIA, who've enabled the front end of the Gen 4 market. We believe the Gen 4 market adoption rate hasn't been as great as it could be, pending when the major semiconductor company will get their Gen 4 chipsets and CPUs into the market. And as far as Gen 5 is concerned, again, I just -- we're going to be prepared for PCIe Gen 5, most certainly. But it's -- why customers should go now instead of going later is don't put yourself at the mercy waiting for PCIe Gen 5 to roll out on major chipsets and platforms if there's any delays, and your competitors have already upgraded to Gen 4, and that's just another year or so of them being able to out-execute you if -- AI and data sets, if you need to move on those quickly.

And from OSS side, we've seen our whole AI transportable market is the power of the data center at the edge. So if the data centers move to Gen 4, our customers immediately want to be on Gen 4. When we move to Gen 5, they want to be immediately on Gen 5. So that's why we work with companies like KIOXIA that keep up with that. So we always have the latest data center product available put out on the edge. So if you wait, you're just going to be delaying, like Matt had talked about.

Okay. Good. Next question, also, Matt, has dual-port high availability been a popular feature in your products?

Yes. I mean, there's only a couple of us in the market, right? If you were to take a look at the analyst information that's out there, storage is the main use for a dual-port drive for high-availability applications where you have multiple paths to the same drive. There's us. There's another SSD vendor and another one just came into the market with the PCIe Gen 3 product. So there's very limited numbers of PCIe dual-port SSDs in the market. Whereas for SaaS, well, even the SaaS market has sort of died down by quite a bit as well. But because SaaS has been a very prevalent technology for the high-availability market, that's continuing to be strong. However, we do see, on the horizon, quite a bit more flash forward systems like the FSA that take advantage of dual port in NVMe as NVMe is a lot faster than SaaS is.

Okay. Good. Question for Jim. Are you doing anything in your system designs to support the FIPS 140-2 in addition to just including the KIOXIA drives?

Yes. So that's a good question on this. Both the data at rest that is available from KIOXIA makes the customer be able to have a choice. So on the One Stop side, we add encryption modules to the Ion Accelerator software, for example, for the SAN and NAS. So we can do the other type of FIPS 140-2 encryption that Matt was discussing when he showed you both models. So it really gives our customers a choice in whatever they have. Because sometimes, they already have a key management requirement or -- and we're just upgrading to a better system with ours. So they can keep that if they're using software-based encryption or move to the KIOXIA in the drive level encryption like we showed in that mobile data center-type system.

Okay. Good. Well, that looks like we've hit all the questions. I'll go ahead and pass it back to you, Jim.

All right. Thanks, Tim, and Matt, thank you. We're going to wrap it up here with a little bit more on the disclosure side, and thank everybody for coming to the webinar.

Thank you, everyone. Have a great day.

Thank you, everyone, for attending. If you have any questions, feel free to you email us after the webinar. Thank you.