3D XPoint in 2025 and How We Got There
3D XPoint has become known as a flash competitor that has actual products, use cases and support. Analysts expect markets for 3D XPoint to rise steadily. Here is a look at the market.
Download the presentation: 3D XPoint in 2025 and How We Got There
00:02 Milind Weling: Alright, so let's get started. We have four panelists as I mentioned, and our first panelist is Mark Webb. He's a Principal Analyst at MKW Ventures Consulting. His focus areas are memory, technology analysis, costs, markets and roadmaps. Previously he worked at Intel for 23 years in manufacturing, process integration and product engineering.
Our second panelist is Chris Petti who has been working in emerging memories on and off since 2000 and is a veteran of Matrix Semiconductor in SanDisk/Western Digital and is now at yet another emerging memory startup.
Last but not the least, there's Mario Laudato. He received his advanced degrees in electrical engineering from Professor Ielmini's group at Politecnico di Milano, focusing his study on emerging NVM and selectors, 3D XPoint memory and neuromorphic computing applications. He's now working...
02:23 MW: Okay, you can go to the next one. So, I just have a couple of files taken from a presentation I just had an hour ago. A couple of things on product forecasts, Optane and eventually Micron 3D XPoint SSDs will differentiate themselves more from the fast NAND products in 2021 plus. Right now, they're kind of in the same fast SSD market. This is based on controller improvements, and the fact that PCIe Gen 4 will allow the chip and cell speed differentiation to sign through.
This article is part of
Flash Memory Summit 2020 Sessions From Day One
So, we're expecting more differentiation from Optane compared to the other fast SSD products. Optane DC Persistent Memory will grow steadily, Gen 1 and Gen 2 DIMMs are continuing to sell and grow steadily. But in my opinion, the future of persistent memory is not on the DRAM bus. The right answer is for multiple sources, growth in easier developments as a new bus. Most people expect to see CXL bus, but it could be others, and in my opinion, this will lead to an inflection point in revenue in the 2023 time frame. If you can click one more time to do the build.
03:44 MW: So, if you look at last year, I had forecast for revenue for the next couple of years. I lowered those a little bit in the short-term and then kept them the same in the long-term. In the short-term again, the takeaway you could have is that the 3D XPoint SSDs haven't taken off quite as fast, and the DIMMs have been a little bit slower than expected. They are still a very large revenue compared to everything else that's out there. And then with Cascade and Cooper Lake share looking at the DIMM attach rate in the average Optane density, we can kind of project the revenue.
The big difference you see is that there's a big jump from 2022 to 2024. That, in my opinion, is alluding to the fact that at that time these next buses will come on board and that'll allow other architecture, so you don't have a completely proprietary DDRT, so called the interface. That'll allow more on the DIMMs or it won't be DIMMs in this case necessarily, but it will be persistent memory.
Next. So, if I look at what 3D XPoint will look like in 2025, 3D XPoint will be the dominant persistent memory technology in 2025. It's already probably the dominant persistent memory by far, but it'll continue to grow.
05:00 MW: I expect 4 1/2 billion gigabytes to be shipping by 2025 annually, greater than $4 billion in revenue from Micron and Intel and any partners they're working with combined. Persistent memory will be accessed via the DRAM bus using proprietary Intel architectures and the CXL bus. RDMA persistent memory or anything over fabric will be available, which will allow more applications.
The fastest SSDs are shipped with 3D XPoint, so this will continue to be, so anybody who's looking for the fastest SSDs, the densities will be something on the 10-terabyte-ish density but they'll be the fastest SSDs will all be on 3D XPoint.
And competition exists, but one of the things is with new markets in my opinion, the competition won't be as much cannibalization or reducing, but it will fuel growth and adoption. One of the number one things people complain about with 3D XPoint today is, "Well, there's only one persistent memory out there." They want other options, even if they choose the 3D XPoint, they still want other options just to look at. And so, I think it's one of the rising tide floats all boats type of thing that'll exist for 3D XPoint. So, there will be competition, but it won't cause 3D XPoint to drop off. That's all I have.
06:37 Weling: All right. Thank you, Mark. The next talk or the next presentation will be by Chris Petti. Chris, please go ahead.
06:48 Chris Petti: Thanks. You can go to the next one. So, I want to talk about first what applications can benefit from 3D XPoint's value proposition, and are these niche, and what is the benchmark? So, this kind of dovetails with what Mark was talking about.
So, what are the value propositions? First, that's the lower latency than flash and the hundreds of nanoseconds to one microsecond. It's got a lower cost per bit than DRAM, somewhere between 30 and 50%, and of course, it's the persistence of the data.
So, I listed four applications here and there may be more, but these are I think, the broadest categories. So, the first is high-performance computing databases, and the benchmarks for that are a large capacity and a low cost of ownership, and a low latency. So right now, that's mostly DRAM, this is what the 3D XPoint DIMM is trying to sell into. So, it is a good fit there. I put as growth, it's stable to declining, although it's kind of what Mark was saying in the previous talk, and if you go by Intel's own talk from the 2020 VLSI Symposia, this application is really the. . . It's not growing so much in share although it may be growing in units.
08:18 CP: Whereas the next application, which is AI for training and for inference, is probably growing more and that I put that arrow down to the client-generated media section plus, plus some of the broadcast data as well. And one thing to note is to take advantage of the data persistence, the software stacks must be rewritten, and encryption is required because it is persistent. So, it's going to stay there so that DRAM doesn't really care about that, but persistent memory will.
So, these are some changes that customers don't have to adopt but they can. But for AI really you need very high bandwidth and low write energy per bit. So right now, the incumbent is the high bandwidth memory DRAM. The 3D XPoint is not the best fit for that because the write bandwidth especially is not high enough for this. This is expanding a lot, obviously this market.
And then the last two are mass storage, the "HP mass storage" is high-performance mass storage which is Optane as Mark was saying, and this is the growth is again what Mark was saying it's . . . Depending on who you talk to, as more adoption comes along, it may be better but it is a good fit for 3D XPoint obviously and it is a good competition to Z-NAND and NAND-based solutions for that but for general mass storage, the lowest cost per bit, and it's not made to fit into that market.
09:52 CP: Okay, you can go to the next one, Millind. So, performance, reliability and scaling concerns. I'll try to go through this quickly. Performance and reliability concerns really are the . . . As I said in the last slide the write energy per bit, and the endurance, there's some question about the endurance of phase-change memory compared to DRAM.
And as far as scaling goes, I divide it up into scaling in the x-y plane and scaling in z. The x-y plane, you do get some advantages because the write currents will decrease, but the voltages will not, so the driver sizes are going to stay roughly constant so as you decrease the CDs, you need to fit all of the drivers underneath the tile, so if you . . . Unless you deal with the decrease in array efficiency. So the way you can do that is just to increase the size of the tile but this will increase the write, this will make the write power consumption worse because you have more of the cells leaking across the tile, and also of course scaling beyond the 20 nanometer node is going to require double patterning so that's going to increase the cost.
11:01 CP: Scaling in the z direction is limited because you have to . . . Adding more decks is going to increase the process costs and it's pretty linear. And also it's going to require more drivers to drive these decks, and they have to fit underneath the array unless you want the array efficiency to crater, so the optimal number of decks is about 4-6 by most calculations.
This is a lot of words, but the question is, can neuromorphic computing leverage the 3D XPoint? And the short answer is not really because there's been a lot of work on phase change for neuromorphic computing mostly from IBM and Macronix and their joint venture. And they've gotten some very good results, I think, but it does require precise control of the storage/compute elements. And that's a little bit difficult to do with the threshold selector, which is the key part of 3D XPoint, it's what makes it 3D.
So most of these applications have been using a select transistor as these pictures imply, so it's a little bit difficult to use the 3D XPoint general architecture for neuromorphic computing, but some of the technological advances that 3D XPoint has made with the phase-change element, that could map into this application. So that's all I have.
12:32 Weling: Thank you, Chris. To keep things going, Mario, please present your slides and go to next slide.
12:40 Mario Laudato: Thanks, Millind. I will focus more on material devices so I will start from the element that we have an offer for unit cell for 3D XPoint is formed from by PCM and OTS devices, as we know from the TechInsights' report on the first Optane products, and for PCM for both PCM and OTS, we have a drift of amorphous phase, and in PCM memory it limits the possibility of multi-level storage in the single cell and adds a lot of design constraints. So, during time, the property of a PCM memory, especially in the amorphous phase tends to change during the time, so if we have a multi-level storage, the possibility to have a multi-level storage is really really limited by the drift of the PCM memory.
13:46 ML: And in addition to that, also the OTS element has intrinsically threshold of the threshold voltage of the selector, so for both the drift is a very, very important component, and it is something that during years, in the last year is very very focused in the development and the tuning of the composition in order to reduce the drift for both elements so the PCM and OTS element. And in addition to that, the leakage versus threshold drift is for, especially for OTS selector is a critical trade-off when we have to select the right composition, the right chalcogenide alloy for the 3D XPoint application.
14:46 ML: So, the second point is that in addition to the material, tuning and screening, also the 3D XPoint as we . . . Also as Chris mentioned before is very challenging on multiple layers and the high process integration complexity, the high number of different edge and alignment, self-alignment step in the process and we have also the viability deck-to-deck in terms of performance due to thermoelectric efforts or other components that are dictated by the fact that we are controlling it with the . . . We have a word line and . . . In common to the two layers.
And the second point is that the actual PVD process at the... It is the actual BKM for 3D XPoint, limits the film homogeneity on large scale, probably also precluding integration of tens of decks and there's the obvious reduction of cost per bit. So, in this sense, there are a lot of challenge also in order to stack, not only in terms of process integration, but also for arrays, a variability within the array. So, next slide, please.
16:22 ML: So, I'm trying to point out the possible future scenarios and solution for this issue. The first one is very interesting that in the last year, one or two year, it's increasing a lot of the number of publications in terms of using the phase-change heterostructure or a superlattices PCM. Here on the top right you can see a paper publishing on Science last year where stacking a single layer in a heterostructure memory, we can have it with a titanium telluride and antimony telluride single layer stacked in the multilayer. We can see that at the amorphous phase, it doesn't show very, very high drift, it tends to be very stable during time.
And the second in the second plot, we can see also that we can have in this way, we can have a multiple states level and we can have probably better . . . The possibility to store within the same cell different higher number of bits. So, it would be probably a breakthrough technology going forward, if it is proved to be sort of a good technology and reputable technology in large scale, because also they demonstrated good endurance and also lower programming power.
18:16 ML: Generally with the superlattices, as being sure that we can have a lower programming power, so and in terms of also reducing the programming power that is generally a critical problem for PCM, we can have improvements and as I said before, also having multiple level states cell, we can also enable potential neuromorphic computing application where we have to have a sort of analog type control of our synapse and also we need a good retention of the different states in order to enable inference or a bad propagation algorithm.
The second future scenario is the 3D XPoint vertical integration. So, at Intermolecular, we have shown in this year that we have developed an OTS, quaternary OTS composition ALD, quaternary OTS composition with the similar performance compared with PVD. And in this way, we are able to integrate the . . . Changing the paradigm of the 3D integration, we can probably try to have more layers, reducing the . . . Higher density, reducing the cost of our technology. So, going forward, having ALD PCM and ALD OTS material develop, we probably also try to integrate and try to solve the problem in terms of high cost of our actual technology, increasing the number of layers having a different scheme. This is what I have.
20:23 Weling: Thank you, Mario. I think in the interest of time, I'd like to go to the questions.
20:27 ML: Yeah.
20:39 Weling: Okay, do any of you have questions? Please post it in the chat, so we can respond. I'm not seeing any questions yet. Elizabeth, do you see any questions?
20:56 Elizabeth: No, I don't right now. I'll let you know if I do.
21:02 Weling: Okay. So maybe to get things going, I have a question for Chris. Is there a niche application or an incumbent memory type that you would think 3D XPoint has the best chance of replacing?
21:24 CP: I think that . . .I don't think that it can replace any of the big memories that we are talking about, it's not really a DRAM replacement for all of DRAM, it's not an HBM replacement. It's kind of like . . . I kind of agree with Mark that as new interfaces come along, new applications can utilize a lot of its benefits, including persistence, which you might think right now is kind of difficult to utilize in the mainstream, in these mainstream applications that we're talking about, but I don't see it replacing anything except for maybe the high-speed flash, which is a Z-NAND category type, type boxes which are . . . Now, Mark knows more about this, it's somewhat neck and neck, I think, but I think it's going to come down to cost and performance there.
22:29 Weling: Yeah. As they say, the replacement to DRAM and NAND are DRAM and NAND respectively.
22:35 CP: Right.
22:35 Weling: That's probably going to stay for...
22:37 MW: There's another panel session going on parallel with this, which is talk about replacing NAND and I go, I answered that last trace of it, I'm not going to take any more questions on replacing NAND because it's practically free compared to other memory.
22:51 Weling: Mark, it's the year 2025, and we are all looking at 3D XPoint having made a lot of progress, so in that year, assume that you're transported to 2025, who do you think would be the competitors to Intel and Micron? I mean, like Optane and QuantX are the incumbent, so to say, do you see the other big tier one memory makers having 3D XPoint or do you see even some tier twos jumping in?
23:30 MW: Well, I think that tier one, if you've been in the memory business for a long time, and certainly the last 15, 20 years, I said, Samsung generally comes after you at some point, no matter what you're trying to do, so you should just be ready for that, and you should be prepared and be ready to fight.
So my comment was, a couple of years ago, I believe, was I believe Samsung, Hynix, other companies, Western Digital has done it, other people have done it, have done research on it, and it's just a matter of . . . I don't think they saw the market, so in my opinion, they have the ability, and maybe Intel and Micron did all the legwork, but they have the ability to jump in there, and is it going to be completely optimized? I don't know.
But again, I can't think of a market where Samsung said, "No, we don't want to be involved in that." They generally go into every single market, so I would expect Samsung to give it a shot, and then I think, again, with these new interfaces, I'm not an architecture person, but I talk to architecture people all the time, and they really think that's going to be a breakthrough because it'll take away the painfulness, if you will, of trying to design architectures in and then you could theoretically say, I'm going to buy half Samsung or half Hynix or half Micron, half Intel, so that's going to be a big influence. So, I would expect all the major companies to go in there.
24:54 Weling: Alright. Mario, if you are with Mark in the year 2025, do you expect . . . I think you touched on it in your presentation, but do you expect some devices or materials that are much more, maybe FAP friendly, less expensive than chalcogenides, to be potentially replacement candidates to . . . They might still operate as phase-change memories but are there some alternatives?
25:30 ML: Yeah, I mean, in the last period, especially on both sides, so that as I mentioned, the PCM and OTS multilayer approach, it's very, very interesting approach, as also AI in Dublin conference, they showed some interesting paper on the OTS multi-layer stacking. I probably, in thermal process probably would be not totally cheap, I think, but yeah, in terms of composition tuning it's still something that the big company, I think that are still working on, trying to fine-tune the composition, optimizing the crystallization speed or the retention, so it's still probably . . .
I know that it's something that started more than 10 years ago, but I think that it's still on PCM and on OTS side, there are still a lot of opportunity in terms of optimization of the process, so in terms of materials also, so . . . And also, as I said, that the leakage versus threshold drift trade-off is still very open, open challenge, so I think that in the next year it would be continuous improvement in that sense for the big player on the field.
27:06 Weling: Thank you, Mario. There is a question from Mike Midy to Mark. Mark, do you think JEDEC NVDIMM-P, assuming that spec releases fairly soon would be a direct competitor to Intel Optane DC DIMM and why?
27:27 MW: Actually, Mike might know more than I do about this, but historically, I think I've said the last couple of years that we were all waiting for NVDIMM-P to be kind of the generic version. I can't mention company names but I know at least two companies were invested very heavily in making this work, and yet it hasn't arrived, so we've become more pessimistic on it, but I think if it arrived and if . . . Depending on where it fits with the other buses, I think definitely it should be out there, we would love to see it, and I think most people would love to see it, but it seems to be very slow at gaining traction.
28:07 Weling: Okay. Maybe a question for Chris, how about the embedded NVM space or L3 cache, do you see XPoint memory moving in that direction just to enlarge it for brand or from a performance and density standpoint, that's not . . . Or cost standpoint, those are not really good areas for 3D XPoint to focus on?
28:36 CP: Right. There's a possibility, again, as a replacement for say, eFlash, which most folks are working on, most founders are working on MRAM for that. So embedded phase-change with the selector with the OTS, could . . . It's more complex than . . . In some ways, the process is more complex than MRAM, but in some ways it's less complex as you don't have these crazy noble metals. So, there may be some advantage to playing this, you can put it above the array, that the thermal budget is not going to kill your CMOS so there's a possibility there, for embedded.
29:29 Weling: Sounds good. I think we are running out of time, I do want to thank everyone. I think your thoughts on where we will be in 2025 are really appreciated, seems like a bright future. Maybe to leave you on a bright note, at least in year 2025 we can bet that we would not be dealing with COVID and we would have the FMS Conference in 2025 will be a live in-person event. With that, thank you so much and I hope to see you soon.
IDC Real-World Applications and Solutions for Persistent Memory