Grey Beards on Systems - 136: Flash Memory Summit 2022 wrap-up with Tom Coughlin, President, Coughlin Assoc.

Episode Date: August 20, 2022

We have known Tom Coughlin (@thomascoughlin), President, Coughlin Associates for a very long time now. He’s been an industry heavyweight almost as long as Ray (maybe even longer). Tom has always been very active in storage media, storage drives, storage systems and memory as well as active in the semiconductor space. All this made him … Continue reading "136: Flash Memory Summit 2022 wrap-up with Tom Coughlin, President, Coughlin Assoc."

Hey everybody, Ray Lucchese here with Keith Townsend. Welcome to another sponsored episode of the Greybeards on Storage podcast, a show where we get Greybeards bloggers together with storage assistant vendors to discuss upcoming products, technologies, and trends affecting the data center today. And now it is my pleasure to introduce an old friend, Thomas Coughlin, President of Coughlin Associates. I saw Tom at the recent Flash Memory Summit in Santa Clara. Tom, why don't you tell us a little bit about yourself and what's new with Flash Memory? Sure, glad to do that. And thank you for inviting me to the show.

So I've been involved in digital storage and memory since for about 40 years, actually over 40 years now. And I've worked at engineer by background. So engineer, engineering manager, and executive at several companies making storage products. Originally, I actually started out in magnetic recording. Yeah, didn't we all? Floppies, yeah, floppies, tapes, hard disk drives. Yeah, yeah, yeah. And about 20 some years ago, I ended up after being laid off from a failing startup company in the Bay Area. I ended up starting my own company. So I'm still based in Silicon Valley. And I have my consulting company that I've run for many years now. I do consulting of various sorts, technical work as well as market and technology analysis.

I also write things. I do a regular column on storage and memory for Forbes.com. I do reports, well-respected reports on digital storage and memory and application topics. I do a report on digital storage and media and entertainment that I've done for several years. And also another one on emerging memory, non-volatile memory technology. So what was your role at Flash Memory Summit? So I was the program chair for the flash memory side okay yeah and i've been involved actually in conferences for for decades i started the

storage visions conference which is a partner show to the consumer electronics show ran that for for many years and also another show focusing on digital storage in professional media and entertainment called Creative Storage. And then I was, go ahead. No, I was going to ask. So Tom, it's a pretty niche storage event. Everyone asks me every year if I'm going, I should go one year. How was the audience after a few years?

So the audience, this show, of course uh last year there was no flash memory summit in 2021 uh there was a virtual event in 2020 the last physical event was uh 2019 and i would say we were about uh 70 70 about 70 percent uh in attendance for what we were in 2019 that's that's very good compared to what's going on the the rest of the industry, right? Exactly. Yeah, I'm hearing the VMware Explorer or VMworld, I'll always call it. I think they're looking at less than half of what they usually pull in. Yeah. Yeah.

Well, I was at the CES show earlier this year, and that was way down both in attendees and exhibitors than it usually is, for example. The other thing I noticed was the floor seemed like it was pretty big and the booths were huge and they were expensive booths. I mean, I remember a couple of years back they were, you know, they were not much at all. Yeah. Keoxia had a two-story booth, for example. Yeah. Yeah. Honest to God. I haven't seen that since VMworld. There were some big, there was a big there, and all the big names were there. We had keynote talks by all but one of the major flash memory manufacturers.

Micron was not there. Right. But they had their own little event that was in the Levi Stadium across the way from the Flash Memory Center. I guess I wasn't invited to that. I'll have to talk to my friends at Micron. You know, I was not at Micron this game because we're going to get into the 200 NAND. Micron is a customer of mine, and I wasn't invited. Yeah, yeah, yeah.

So I think they pre-announced that they had reached 200 layer 3D NAND kind of thing. Wasn't that the case? Yeah, I think they're 232 232 they announced that the week before the flash memory and then at the flash memory yeah solidime uh announced uh i think it was 238 uh layers solidiner sk hynix announced 238 layers so they got six layers over on Micron. If layers is the only thing that matters, which it is. Yeah, exactly. I thought Kioxia also announced that they were coming out with something. Well, Kioxia and Western Digital both talked about

what's really involved in scaling flash memory. And only one of the factors is what they call the vertical scaling. The other factors are the lateral scaling. You know, they've got these holes that go down into the silicon. And the size of those holes also affects the density. Also what they call logical scaling. That's how many bits per cell. And there's even talk of PLC.

Yeah, I saw PLC mentioned on one of the slides. It was crazy. And one mentioned six layers. One of the companies mentioned six layers. And then the other one is architectural scaling. And that's things like putting the peripheral logic that supports the NAND flash under the NAND. But then there are also various talk about doing wafer stacking, sorry, die stacking, or even wafer stacking. And in fact, YMTC, that was their big claim at the conference,

was that they were doing their third generation of wafer stacking where they build their logic wafers separate from the NAND flash layers and then they bond them together. Tom, this is something that I've been, I understand like the significance of the manufacturing improvements and processes to get to over 200 layers. What I'm not quite grokking, like I'm a processor systems guy, so I get like when someone says this is built on 8 nanometer process

versus a 10 nanometer process versus a 5, I know not to necessarily equate that to better performance, but I know because of the smaller processes that I'm going to get overall better energy consumption more more logic more functionality more a lot yeah more functional more cores even yeah yeah even more cores so help me kind of connect the dots on how is how this is

going to help in the packaging of nan and what that looks like how that how is is going to help in the packaging of NAND and what that looks like? How is this going to change my everyday? Well, I mean, for NAND, and increasingly SSDs right now are NVMe-based, it means that I can have more memory, more storage available to support, and especially NAND Flash and data center applications is primary storage. And so it supports the ongoing activities that are that are being done so a lot of we're generating enormous amounts of data um and when you do like artificial intel machine learning training and things like that there's huge amounts of data are required to be able to do that and

you need to be able to store that on a fast memory technology, and that's where the NAND flash fits in. But another big thing that was going on there too was CXL and some interesting things going on with that CXL interface, which is a true memory interface also based on PCIe, just like NVMe. Right. Before we get off that 200-layer thing, doesn't this mean that we'll get, I don't know, 20-terabyte SSDs,

40-terabyte SSDs kind of thing? Is that where this ends up? It could be. Actually, if you want to go way out there, Samsung was talking about a 32- die structure uh that within 10 years 32 man chips effectively stacked on top of one another is that what we're talking about exactly okay yes in 10 years they said they could do a petabyte with that a petabyte and an ssd a petabyte and a 32 stacked die ss uh device you know so i don't know if i'm going to call that an

ssd it's an ssd of some kind yeah yeah i'm i'm i'm really excited with nvme especially cxl i'm excited to look at what this packaging is going to look like i don't know if we're have we had any talk of what how this stuff is actually going to physically look and how it's going to interface with systems when pooled behind a CXL switch and then share it between CPUs, server CPUs. And so that was the original push on that. Now, the other thing about it is it allowed you to have different types of memory. They didn't all have the same performance. And part of what this was about was a push to use what's often called storage class memory but the particular

example was a 3d crosspoint or intel's opt-in combined with dram and intel was a big promoter of cxl as a consequence of that so but with intel's decommit on optane memory they said they were winding it down yeah now at the Flash Memory Summit, there were a lot I mean, CXL was one of the hottest topics there. And there was a lot of folks that were talking about Flash Memory with CXL interfaces and building CXL devices with Flash Memory, sort of to fill that hole that Optane was hoping to fill.

And CXL is going to be the basis for memory in a lot of upcoming, especially on the server and enterprise side systems coming up. The other thing I thought that I saw, XL, Flash, and a couple of the others, there have been all of these in the past mentions of NAND devices that were storage class memory wannabe kinds of things. But those things seem to be more sophisticated, more maturing nowadays. Is that what you're seeing as well? Yes.

I see that people were really looking at, you know, for instance, on the CXL side of combining NAND flash in a CXL device. Now, maybe it's behind a bunch of DRAM, but it is bringing, you know, NAND flash and maybe it's single level cell or MLC. It's probably not the high density because then you get, you know, that slows down your performance, but they could use some form of NAND flash in order to provide a non-volatile memory that supplements DRAM. DRAM is still pretty expensive. The price has not gone down on DRAM.

That was what Optane was hoping to be able to provide, is a lower-cost memory that could supplement DRAM. But Flash looks like it's going to be expensive. So what's your take on the Optane thing? We've all had high hopes for 3DX, Crosspoint, all that stuff we thought was going to be taking over the world of storage here. Yeah. Intel sent me a terabyte of the stuff to test in my lab.

Did they really? I wish I could have kept it. It was awesome. I'm like, wow, this stuff is awesome. It really was. I found some really weird stuff. I found that you had to oversubscribe at a minimum four to one. So I had to take RAM out of my system to get it to work in the system.

So it was well engineered stuff. It was really, really cool. It was pretty much seamless in my vSphere environment. I really had high hopes. Yeah, I did too. And I think there were,

the big issue they had is that they had to sell it for less than DRAM, but they couldn't make it for less than DRAM. And so they were hoping that if they subsidized it, and demand could develop, that they would get their manufacturing volume up enough that their costs would go down and they never reached the level of volume that they needed to reach a break even estimates uh we've got estimates um jim had you and i from uh the report we do on emerging memories that over seven billion dollars was spent by intel um you know and supporting uh the optane launch you know uh and so when they said they're winding down, actually, I went and talked to them. The Intel had a booth at the Flash Memory Summit, even though they're not officially in memory anymore. But a lot of what it was about is to talk to people

about what's happening with Optane. So I asked them, so what does winding down mean? They said, well, we're still supporting our customers. We're still meeting, you know, demand requirements and they're, and they're qualifying their third generation opting memory devices, you know, with customers. So, you know, this is, it's from that's from what they said, it sounded like it's going to be, you know, it's not a sudden, you know, we're just going to drop everything, but they're going to be supporting their customers and trying to build qualifications. You know, it's even possible somebody might buy that, although it's hard, you know, if it never reached the point where it was even break even on manufacturing.

Yeah. It's hard to see that happening. Yeah, yeah, yeah, yeah. And we think Micron would have picked it back up. You know, they know more about it than anybody, right? I mean, they do, but they they're the ones that dropped out first last year. Yeah, yeah, yeah, yeah. Yeah, I was mad.

That was the catalyst losing, what, at least half of their manufacturing process? Micron built half of the stuff, if I remember correctly. Or more, I think. They built the bulk of the- Yeah, so that- Yeah. Huh, huh. They built the bulk of the Optane memory devices.

Right, right, right. Well, Flash, go ahead. You know, it's a tough thing because if me advising customers, I would still advise them to buy it because if you can get it at a good price, it is so much cheaper than DRAM. And I'm hoping that there's somewhat of a fire sale. Pick it up while you can. Pick it up while you can.

The systems that they typically go in, the lifespan of those systems are not typically long because you need cutting, trying to get the best price per performance mix and you get rid of the overall, and you're going to get rid of those systems in three to four years anyway so you know why not yeah yeah yeah it's i mean it really did provide

a lower cost uh you know boosted up your memory at a lower cost than dram that's for sure and they're uh and you know i guess that leads to the question, what's the replacement? Well, first of all, I think, and I wrote a blog on this. I said the gifts of Optane for my Forbes blog. And I was talking about, so the development of Optane played a role in, for instance, SNIA's non-volatile programming model that they developed. Although it was around before Optane was announced in 2015, but it still, I think, it impacted an awful lot in terms of what that was about.

The big thing it did, I think, is it stimulated the development of CXL, which is built around the idea of I could have memory that's available to processors that could be different types of memory or different kinds of performance properties. And I didn't have to build, you know, it could just be handled by that, by the fabric, by the, the, that CXL fabric. And, and so I think that, that was one of the big gifts essentially of Optane, you know, to the industry with CXL. It seemed to me that Optane also kind of caused the creation of NVMe as a protocol.

I mean, up to that, SCSI was fine and all that. But now, with this stuff that's almost memory class access, you needed something that bypassed all that overhead and stuff like that. So NVMe came into existence, in my mind, primarily because of Optane. Now all the SSDsds in the world are using it and everybody else but including cxl but you know so that's an interesting question that uh that brings up an interesting question how much of the memory controller stuff out of the xeon scalable stuff made it into cxl as a result, if any, of that stuff. Because that was like the secret sauce for Optane plus Xeon is the memory controller

and getting the hit rates right, the caching hit rates right, et cetera. I think there was probably a lot of that that went into CXL in order to be able to provide the coherence and the other properties that they needed to have. But another interesting thing, by the way, at the Flash Memory Summit is not only CXL what it is, but it's becoming more. In particular, there is a group promoted by IBM. There's a microchip called OpenCAPI, right? And OpenCAPI had an initiative that was called the Open Memory Interface, which was a direct memory interface, but it provided the capability that approaches that of high bandwidth memory so um that became that was announced to be that's going to become part of the cxl initiative so open capi and omi that open memory interface are now part of cxl so you may see challenges or at least

options to ddr memory with this omi interface as a part of CXL. So I call that near memory. It's directly connected to the CPU. And then far memory with somewhat more latency is the CXL. So we may have whole new technologies that can affect the architectures that would be promoted by the CXL initiative, including near and far. So you think a DDR is going to go away? I mean, God, they just came out with 6X or something like that, right?

No, no, it's not going to go away. But I think system architecture is going to have more options. If DDR, if a limited amount of memory is fine, DDR works well. But if you want an awful lot of near memory, then something like OMI may be a really good option. The other thing I saw was there was a

software-defined flash? Software-embedded flash? Somebody was talking about a software control layer for flash. I mean, it always was right within the SSD software control, but this was exposing it to the host, right? Yes, there was software-defined

stuff that's been out there for a while. Let's see. Solidigm said something they called Synergy that they were offering with their ssds that would provide includes a storage let's see software running on a host machine called solidime synergy includes they're offering for instance on a client device that they're p41 but i believe they're also we're doing things with that, with enterprise and data center SSDs as well. And they said, actually, that's another one of those gifts from Optane that some of the capabilities were actually taken from the Optane project. Of course, Solidigm used to be Intel's NAND flash memory.

Right, right, right. NAND flash memory. But there are a number of other things where people are talking about new ways of being able to control what's going on inside of the SSD. Every company had their own angle that they were talking about. In our business, Pure and IBM and Hitachi had always done their own flash module kinds of things. And the contention was by having a more global view of the flash, they could manage it better. They have less write amplification, less over-provisioning, that sort of thing. But these guys are making it available so that just about any operating system in the world, any hypervisor, any of those guys could just pile on a bunch of NAND and let it rip, right? They handle the whole flash translation, the page management, garbage collection, all that stuff could be done almost at the host level.

Yeah, there's a lot of interesting options, and it's interesting to see how things go there. For instance, IBM has their Flash core module right right um you know which includes a whole bunch of mram cache as and that's part of its secret of performance is that it can hold a lot of metadata and a lot of other stuff in mram which is a lot faster than the nan flash and has some more symmetrical read write speeds they can even do write caching. So they use QLC Flash, and they claim that by using the MRAM as a cache that they can get endurance that makes this viable

as enterprise-level storage. So there's some fascinating things people are doing out there. But you're right. Some of this, a lot of the SSD guys are trying to democratize some of these control capabilities and, you know, whether they can do it, get every kind of feature that the system guys could want remains to be seen. But there are some interesting angles. There are fascinating things that are going on in terms of architectures and products out there right now and including software yeah as well so tom one of the things that i'm curious about me and ray are uh enterprise i.t focused folks but flash memory summary summary obviously is uh looking at the entire market

as we're looking at technologies such as cxL, 200 plus layers, what are some of the use cases that came up and even just what the gap that Optane tried to fill? What are some of the use cases that we're seeing outside of like enterprise IT for some of these advanced technology, memory technologies? Well, first of all, one of the big things that Optane, you know, use cases that they were successful in was databases. You know, if I can put all my database into this Optane memory, I got really good performance. They were doing, you know. Online stuff, right?

Yeah, exactly. So remember the people they'd bring in to give talks, the database guys love this, you know and stuff right yeah exactly so remember the people they bring in to give talks the database guys love this you know um but uh the other thing that's going to is driving memory in general is going to be um artificial intelligence machine learning in particular both the training but then also the inference you know the actual use of the models and so that's going to be driving a lot of these memory architectures. And the other thing, of course, that's driving the big data centers is they're always trying to get the most effective use they can

of the resources they've got. And so the idea of doing disaggregation and aggregation of resources, creating virtual machines, spinning them up, spinning them down, virtual machines, containers, spinning them up, spinning them down as needed is critical to them. And some of these new capabilities are important in enabling that. For instance, that's what CXL is going to be part of that. And there's some of the talks, the controller guys, they were talking about CXL combined with NVMe over Fabric that together they give you this capability of pooling and more efficient management

of both storage and memory resources. And I think Marvell was even making some hints that maybe you could tie these together in some way, a fascinating way. And that could span all the way with that. And with, for instance, NVMe native hard drives, you could go all the way from hard drives to the most sophisticated memory technologies that are out there. So you're saying CXL kind of could be similar to NVMe over, could be attached via something

like NVMe or Fabric. So I could have my shared memory tier sitting out there and have a couple of petabytes of memory and have it shared across 100 servers or something like that? They had this slide there where they're talking about composable infrastructure, and then it points to NVMe over Fabric and CXL. And they show sort of these paths going up to the server. They don't go directly from the NVMe over Fabric to CXL, but they're working together through the servers

and through the stuff that's connected then to them, either an Ethernet switch or a CXL switch. So you've got a storage pool, which would be a heterogeneous storage, can include hard disk drives all the way through to accelerators like DPUs under the Ethernet switch is NVMe over Fabric. And then on the CXL side, you've got expanders, accelerators, DPUs, FPGAs, GPUs, and all kinds of SoC devices that would be part of that. And they're sort of both serving whatever functions that are needed for the servers to be able to do the stuff they're doing.

But I think a lot of this is being driven by we're just getting. Yeah, go ahead. I was going to say, I've been doing a lot of sponsored work with Micron and just kind of normalizing some of this for the general, you know, enterprise architect audience. And that's been the talk track, kind of that, you know, HPE's machine of five years ago, circa five years ago, kind of becoming real memory driven compute. And I think what's really changed is these use cases that you've talked about around AI, ML, database, this need to compose systems that are way bigger than what you can do in a single node that has to expand multiple systems, but you have latency problems between memory, et cetera, et cetera. Somatic compute is really, really hard, but CXL makes it possible.

So when you're thinking about being able to give a GPU, you know, let's say a reasonable amount of 64 terabytes of memory, that's, you know, that's, that's, that's game changing for AI ML inference. Yeah. And these technologies are going to make that possible. And to do it, and even to do it without the high bandwidth memory, that's where you might be able to, for instance, that OMI approach, which is now part of CXL, I can get an awful lot of near memory at near HBM speeds, but much higher capacities than I could get with either HBM because it's done with wafer stacking or with DDR. So it's, or I can get the DDR kind of capacities, but I can get them at HBM speeds, which is really. Yeah, I had watched this.

It was actually on the consumer side. I watched a video. This consumer YouTuber demoed a AMD graphics card from a few years ago that had four terabytes of SSD connected directly to the memory, to the adapter, and they were showing the improvements in AI and ML just from that simple hack. So I can imagine the performance that we'd get when you couple this with a true enterprise-class solution. I guess the question is, do you think, today they've got like 64 gig and 80 gig GPUs. Do you think they're memory constrained? Oh, I think they're memory constrained? Oh, I know they're memory constrained.

Intel folks, here's an example. I'll send you the link in the show notes. Yeah, yeah. Ray, there was a Tech Field Day event where Intel came and demoed Optane challenge between basically Optane and NVIDIA for search. And the Optane system smoked the GPU system. For search? Just on optimizing for memory. Yes, for search.

Yeah, but search and GPU stuff is not exactly – well, I guess they're both scalable compute-intensive activities, maybe. I don't know. I could see that for search, maybe, because the bigger the index, the better the world. You know, the faster you can access it. But we're talking, you know, bring in a byte of data, do some computation on it, update some number here and do that, you know, 10,000 times at the same instant in time. You know, these guys got a lot. here and do that 10,000 times at the same instant in time. You know, these guys got a lot.

Yeah, I'm a huge proponent of getting as much RAM on a GPU as possible. Because I think, I mean, what's the difference? If I can put a whole ML cycle and data set directly onto DDR on a GPU, you're telling me that's not going to significantly outperform something that has to go over the bus to get the, the data, the ML data. I mean, it's, the question is how many, how many compute engines, I mean, so a 4,000 compute engines, all right, maybe 8,000 compute engines on a GPU and maybe those things double over time, but I don't know, you know, 80 computer engines on a GPU. And maybe those things double over time, but I don't know.

80 gigabytes seems like pretty much enough storage to keep those guys fairly busy. So another thing to think about is that this pooling and building these fabrics for memory and for storage, for that matter, allows you also to put in these specialized processors to offload tasks from the CPU or a GPU. And I think we're going to see a lot more of those, a lot more DPUs, a lot more specialized devices that do functions that give you, you know, either take work away from the CPU so it can, or GPU, so they can focus on what they're trying to do, but also provide new capabilities. For instance, IBM, their Flash Core module, they actually have very fast real-time compression decompression they're doing apparently. I understand compression decompression.

You put that on an FPGA, you put that on an SSD, that's fine. That's easy. That's not a DPU. That's not multiple ARM processors sitting out there, RISC-V processors sitting out there running containers that you load from an operating system. That is true. guys are doing in the hypervisor and what uh what aws has done of taking these menial tasks that the cpu or gpu shouldn't be doing or are less way less efficient at and moving that off of the gpu into some near memory uh capability that will inc incredibly increase performance versus what I'm saying, Ray, which is to put more memory directly on a GPU.

What if we just offload some of that overhead and let the GPUs do what they're best at doing? Yeah, but that GPU direct and all that stuff, these guys are trying to get the stuff to the GPUs as fast as they can. You know, it's not like there's a... Well, what if you don't have to get it to the GPU is the question. What if you can do it outside of the GPU? Yeah. Well, and what if this internalized, you know,

network processing actually helps you to get it? I mean, they're even talking about storage devices that are a lot smarter, even know more about the content they've got on them. So what if you could do a search on individual storage devices? The whole compute storage, we've been talking about this thing for now three or four years, Tom. Three or four years? No, no, no. Three or four decades, I think. No, I understand. I did it back in the 90s, okay? Don't talk to me about this stuff.

But the thing is... It's rare when networking is ahead of storage. Yeah. Networking is definitely ahead of storage in this. I think this is... It has... Networking is... Come on, I can run NSX basically on a switch now.

Networking is... Networking is being able to centralize a lot of these functions a lot better than storage has. I think that functionality, these basic computer science concepts are going to catch up. You know how much code runs on storage devices, Keith?

I mean, we're talking millions of lines of code, okay? And networking has the same sort of, you know, NOS and all that stuff, millions of lines of code. I understand that. It's just smarts out there in the end. The question is, each one of these things is a proprietary environment. It's not like it's a hypervisor.

It's not like it's an operating system. It's not like it's, you know, I can run Ray's application sitting out on a DPU. That's what they're doing. So they're getting there. The NSX. there the nsx so uh as you uh as you able to run firewall functions on uh merchant silicon or even commodity silicon and you're taking that capability or need away from the gpu or cpu on the hypervisor is aws's margin i can run firewalls on my Comcast modem out here. It's nothing.

Not at petabyte. Not at petabyte. Okay, yeah. There's a whole different. We're talking about enterprise class. All right, all right. And now the hyper. So once you're able to take that functionality

and you're able to, I think Martin Cassaro is, I forget who it was, but they talked about what happens when CPUs, the cost of CPUs reaches zero. And you have these ARM processors running on storage on SSDs. Someone's going to commoditize that. I just believe that's true. What's missing from that environment is a software-enabled framework that's universal.

I mean, if I can run a Docker container on my storage, fine. If I can have the control that Docker gives for that functionality running outboard, that's great. The problem is most of these things are very special-purpose processors or very special-purpose logic. Yeah, I can run encryption or I can run compression or I can run, you know, buffering or something like that or caching, but they're not letting me run a container. So this is what I said. Not yet. This is not yet. And this is why I said networking is further ahead. There's P4 for networking, which allows that to happen. That you can have the disparate types of processors, et cetera, but you have a single language in P4 to allow you to describe exactly what you're

talking about, Ray. This is going to happen. It's got to be a container. I got to be able to write it in Python. I got to be able to run it on my Linux system and then run it on my storage and run it on my network. I don't know if you have to run it in Python. I think you have to be able to run it in a commonly accepted language. Like what the networking folks are doing are not containers. They're doing something written specifically, a language specifically for networking, P4, and that makes it portable. You need that portability.

It doesn't have to be a container that's general purpose. It just has to be widely accepted. And provide an advantage, right? It's got to be testable on my system. I've got to be able to run it on my laptop. I've run it on my server. That doesn't have to be Python. That can be a Python-like language. It just needs to be, you need to be able to virtualize it and containerize it and move it. It needs to be portable, not necessarily what we use today.

I'd like to see this example of attaching an SSD to a GPU directly and it speeds up some process. Other than search. I'll send you a link to that video. Other than search, this was just, again, this is a consumer-level YouTube channel. And I was surprised to stumble along it. But it was really, really clever. I got friends in the crypto business that might want something like this. To do some mining.

Yeah, exactly. If it helps. I don't know how profitable that is anymore. Yeah, I know. They're all... Well, speaking of that, there were talks at the Flash Memory Summit on time and space-based cryptocurrencies,

in particular, Chia. There was people from Chia. And this actually uses storage. Yeah, I know. I've got a friend of mine that's got a petal by a Chia running. Yeah, and the whole idea behind that is it doesn't take as much processing. It takes space.

It doesn't take as much energy either. It doesn't take as much energy as a consequence, yeah. So you generate your field with SSDs, which is fairly intensive. But then you let them sit on hard disk drives, and you look for matches, and someone else generates the same thing, then you get your coin. I'm trying to convince my friend

to use MAID or something like that so they can actually power it down unless it's actually usable. You need it, but I don't think he's convinced yet. Now, another thing at the Flash Memory Summit was that we've had a whole session on DNA storage. And that's interesting technology probably five years out or more before something happens. I think it's got to be at least a decade time. Come on, it operates on the order of hours, right? I mean, if it takes me an hour to write a byte of data, it's not going to happen. Well, except that there's a lot of stuff

because of the biomedical applications that's speeding up the ability to read and write genetic information. So there's a whole different driver of ecosystem for that, you're saying. Exactly. Exactly. Yeah. Yeah.

So that has to be taken into account is that this is storing what we are. It's got to be milliseconds, Tom. If it can get down to read and write in milliseconds, talk to me. But it's not there yet. It's orders of magnitude, maybe four or five away from that. What's interesting is there actually are silicon-based devices that people are building for doing DNA storage, which makes the possibility of doing something quickly conceivable. It's a ways off. And what we're probably looking at is competition to tape and

optical, you know, for archiving, archiving applications. It's not going to be. These things don't go away, Tom. They just, they just kind of move their niche down or up or over or something like that. Yeah. So performance SSDs, you know, came out and came out and then killed off performance disk and just moved down to a level, you know, near, near line and stuff like that. These things never go away, Tom. They just move. It was, and that move down to a level near line and stuff like that. These things never go away, Tom. They just move. And that just talks to you about the whole value of information in the modern age is that all these different ways to store information that do trade off between cost and performance.

Information is what drives our modern industry, is what enables a lot of things to happen. And so all of these things continue to exist because they're useful to some people. I was in one, I'm going to write a blog post about this. I was in one session, keynote session, and the gentleman there said, you know, we're creating, you know, literally 22.7 exabytes of data a year, but only 10% of that is being stored. Because mostly it's data that needs to be processed real time to be useful. For instance, I've got sensors on a car. The car has to decide if I'm going to hit something or not. If that data needs to be processed in real time to be useful, it can be processed

offline and still be useful. The data is data that can be used. I don't care if it's real time and thrown away. If it's real time and thrown away, we're missing something. We as an industry are not doing what we need to do. Well, but think of it. I've got cameras on my car. My car is trying to make a decision. Am I going off my lane or am I going to hit somebody and should I apply the brakes? That's got to be processed locally. It's got to be processed fast, but it's not necessarily going to be kept long-term. But the results may be kept long-term.

Tom, don't you think using that data to better fine-tune the AI algorithms that are driving your car, don't you think that, even if it only means personalizing it to Tom the driver, that sort of stuff. So what we need is real-time AI training capabilities that are built into these devices, which probably means a lot more memory and storage. We all bring it back to storage and memory. Okay, okay. I see where you guys

are going with this. So Keith, give me that link. I will send it to you because it's a YouTube. Yeah. So Keith, any last questions for Tom before we close? You know, one last question. You're a standards guy. Can you talk a little bit about how we're driving towards kind of this nirvana that Ray would like to see of just standards around not just the storage interfaces, but the underlying compute that's driving it? Oh, that's an interesting one, too, because another thing that Debendra Shah Das Sharma from Intel was giving a talk on CXL. And CXL, as well as something called the UCIE, which is this chiplet interconnect technology, both, again, use PCIE.

PCIE is kind of becoming the universalplets, but also driving into what they call heterogeneous integration, where they can make stacks of chips and even go into things that are like totally a 3D stack structure with vias, very thin wafers with back connections, vias to go through and just do amazing things in much, much smaller spaces, allowing us to make, hopefully, standardized, standards-based technology that would underlie everything all the way from consumer devices to enterprise and data center applications. So I think there's some fascinating things that are going on that are going to be fundamentally driven a lot by the slowing down of Moore's Law. It's driven new forms of creativity and how we can continue the pace of increased processing capabilities. So something like this would be multiple, let's say, CPUs with multiple cores each kind of either attached in a 2D mesh or something like that, or even a 3D and a 2D mesh kind of thing all connected through UEI?

That's what's possible. That's possible. possible what's driving it right now is uh that uh the uh finer lithographic features say five nanometers or smaller are getting incredibly expensive the euv lithography equipment to make these things is over you know so it must be like 150 million dollars to buy one of these things and so you got one in your lap oh yeah exactly yeah so the idea is that I... Oh, yeah, exactly. So the idea is... It's sitting right under my Aletra storage array, yes. So focus the use of that technology on logic, for example, which could benefit most from it.

And then, for instance, my memory may be running at a more relaxed lithography, less expensive to make. It's a little separate chip, but it has a very, very fast interconnect. Very local, very fast interconnect between that and the processor chip. Something like SRAM or something like that? Could be. Could be SRAM, could be MRAM, could be all manner of different things you could potentially do with this. And I can, they're building standards so I can take these chiplets

that come from different manufacturers, put them together, and they will function together. That's the ideal behind it, which if you can do that, it creates this underlying structure that would go into all of our IT equipment, into consumer devices, and everything that would, again, enhance our ability to store and process information and do it at much more localized levels, for example. That's amazing. It's like taking a server and shrinking it down to a chip. Well, I love this idea of disaggregating and then composability.

And that's kind of, we're taking that to another level. Yeah, yeah, yeah. Bizarre, bizarre. All right. Tom, is there anything you'd like to say to our listening audience before we close? Oh, sure. Well, it's glad to be able to talk to you folks. I hope you keep in mind the Flash Memory Summit in 2023. We're going to start doing our call for papers in January of 2020, early January 2023.

It's going to be, again again at the Santa Clara Convention Center. I believe the dates are going to be August 8th through the 11th. And also I've been an IEEE volunteer and I'm running for IEEE president this year. And for any of you that are IEEE members, I urge you to go out and vote. Voting starts today, August 15th. Oh God. Well, good for you. Good for you.

Well, that's it for now. Bye, Tom. Thank you very much, God. Well, good for you. Good for you. Well, that's it for now. Bye, Tom. Thank you very much, folks. Good to talk to you. Bye, Keith. Until next time. Next time, we will talk to the most system storage technology person. Any questions you want us to ask, please let us know. And if you enjoy our podcast,

tell your friends about it. Please review us on Apple Podcasts, Google Play, and Spotify as this will help get the word out.