Download the presentation: 3D XPoint Optane Memory Markets: Bits, Revenue, Costs
00:15 Mark Webb: Alright, my name is Mark Webb from MKW Ventures Consulting, and I'll be talking about 3D XPoint Optane media memory, markets, bits, revenue, costs, and application primarily for data center and what we can do with these new architectures. We'll talk about 3D XPoint Optane history and background, 3D XPoint products, models for the technology, Gen 1, Gen 2, estimated performance metrics in 2020, bits in revenue, we'll look at cost versus other memory technologies, and then finally give a summary of where we are.
So, what is 3D XPoint? And again, I've done lots of details on this before, Micron and Intel tech . . . Micron, and a technology that was announced in July 2015. As a disclaimer, Intel doesn't have 3D XPoint anymore, it's called Intel Optane memory media. So, that's what they refer to it internally. It is a phase-change memory technology, in a cross-point array, with about a 4 F2 cell, which is then stacked. The lithography is about 20 nanometer lithography.
01:33 MW: The key takeaway for this is it's faster than NAND, slower than DRAM, cheaper than DRAM, more expensive than NAND, so it's not replacing either technology, but it's something that fits in-between with certain aspects. What's new this year is the second generation includes four decks, I think they're called, by Intel, or layers, so the first generation was two layers. So, if we look at new memory technologies compared, I'm just looking at a couple of them here, the key takeaway, that change . . . that affects 3D XPoint is we have all the technologies that are out there, but the key is 3D XPoint is HVM ready. It's been running in production, multiple applications, so we have a lot of volume line.
02:22 MW: Everything else, it's kind of in-between on latency, density, and cost, but overall, the big aspect is, compared to other memories is, it's HVM-ready, it's running in volume. So, if you look at a simple architecture overview, this is a slight update to a slide I shared before. One of the things is, 3D XPoint is not three years in the making; it's not five years in the making; it's really not even 10, 15 years in the making; it's been around for 20 years. And if you look at the history, before it went into stealth mode and kind of disappeared nine years ago, Greg Atwood from Micron Technology published at FMS 2011, showed basically what the cross-point technology would look like. There's papers in 2001, 2005, and 2009 that were also published.
This article is part of
Then you have us -- from the middle here, you can see that this is a cartoon that Intel has -- and then you see there's been speculation over the years on the technology, and the one big update was at Intel Architecture Day in 2020. They announced the second generation as being four deck, this is something we actually predicted four years ago, saying it was going to be a four-deck technology, they showed it, and then this is from a foil that Intel presented at the Intel Architecture Day. So, if we look at a product update for 2020, Intel Gen 1 products, they have the fastest SSD with the best SSD endurance, selling millions of units. So, a big application of it is really in SSDs.
04:05 MW: The DIMMs were launched in April of 2019. The sales have been growing steadily, still a relatively low attach rate, and probably lower than Intel would wish, but it is steadily growing. Right now, less than 5% of servers sold today have Optane DIMMs in them, but then again, that is still a significant amount of volume and Intel expects it to grow. There's more Optane or 3D XPoint bits are sold annually than all other emerging memories combined. Intel Gen 2, and I'll talk about this star in a second, products are coming out. So, the cell chip speed and endurance are not significantly changed on the Gen 2 chip. We expect the endurance targets to be about the same, we expect the speed to be about the same, even if it has four layers instead of two.
04:57 MW: The DIMM speed has increased due to a controller change and the bus speed. Gen 2, which is Barlow Pass -- or I refer to it as blue because it has the blue heat spreader on it -- is available with the newest CPUs that Intel's announced. But again, the big in-speed improvement is really from the bus speed, which is due to the controller of the bus. The SSD speed for Gen 2 products is greatly increased due to PCIe Gen 4 configuration and controller.
So, again, what we're seeing is, in my opinion, basically the same type of speed of memory and endurance, but the controllers have improved. And we'll talk about the implications of that. Micron products will ramp on Gen 2 but are still in development. So, there is a product that Micron announced, it's really going to be more of a development vehicle, but we expect revenue for Micron, measurable revenue for Micron, in the 2022 time frame. Now, the definition I said might be complex, if a Gen 2 product doesn't have a Gen 2 chip, what would you call it at that point? And we're still speculating on that.
06:07 MW: Product forecast. So, Optane and eventually Micron 3D XPoint SSDs will differentiate themselves more from fast NAND products in 2021. So, if you want to get the fastest SSD today, you can get what they call fast NAND or low-latency NAND SSDs. The difference is we think that the 3D XPoint SSDs will further pull away from this because the controller improvements are made, and the PCIe Gen 4 allows the chip and cell speed differentiation to shine through, whereas when it was on Gen 3 with older controllers, it was kind of held back, if you will.
The Optane DC persistent memory will grow steadily, Gen 1 and Gen 2 DIMMs will continue to sell and grow steadily, but in my opinion, the future persistent memory is not on the DRAM bus. The right answer for having multiple sources, growth and easier development is a new bus, CXL or it could be others as well. This will lead to an inflection point in the 2023 time frame where we see adoption of persistent memory really take off.
Technology update for 2020. We had Gen 1 chips for the first SSDs, DIMMs and memory modules. That technology was a 128 gigabit device, 20 nanometer. It was about a 200 millimeter Di size, two-layer stack and, according to our estimates, the endurance grew to over 200,000 cycle capability. It started out much lower than that.
07:44 MW: We now have Gen 2, an upcoming Intel PCIe Gen 4 SSDs, which will be announced and released in late 2020. We predicted this before to be a four-layer, 256 gigabit chip, similar cell chip performance as previous technology, but again, higher density, and at the time we're modeling 35% lower cost. Then what happened is Intel announced Barlow Pass, it's the Intel persistent memory module or persistent memory second-gen DIMMS. The shocking item was that there was no change in the DIMM density. Gen 1 was released at 128 gigabit, or sorry, gigabyte, 256 and 512, and the second gen was released at exactly the same density. So, the question is why?
Well, after discussing with some people, the key issue is that customers want lower density products. They're just implementing it now, and if you compare it to some of the other NVDIMMs which are much lower density, customers are not ready for 1 terabyte of memory, that's not a high-volume decision or a product that people want. One online publication stated that this product will actually be a Gen 1 chip in a Gen 2 module.
09:01 MW: Again, we haven't seen the teardowns on this -- we're guesstimating something in-between. My prediction is not quite Gen 2 chips are going to end up in some Gen 2 DIMMs, and we'll see the details soon when we get the teardowns on those devices, so I expect when those come out we'll publish a note on it explaining exactly what's going on and how they're managing those, and the key reason why is because the higher density is not necessarily needed on DIMMs.
3D XPoint capacity and manufacturing. 3D XPoint is manufactured only at the IMFT facility. Actually, I should change the name on that, that is the Micron Utah facility, Micron Lehi in Lehi, Utah. Intel, Micron have a 2020 supply agreement, which they agreed to earlier in the year to supply Intel's needs for the future, so they broke up the joint venture. But now they have a supply agreement going forward, so Intel can get parts from the Lehi facility for the foreseeable future. Micron is still using minimal capacity and as minimal revenue other than Intel in 2020, as I mentioned before, the X100 SSD is more of a development vehicle for future products. That's fine, everybody needs to start off and do the development, and we're expecting revenue from Micron in the 2022 time frame.
10:23 MW: Looking at some financial records, estimates of volume shipped, the Lehi factory is not a limiter in the near future for Intel's needs or anybody else's. No new capacity is needed. Last year, we forecast what the capacity of the factory is, saying that we had Gen 2 and Gen 1 capacity were updated that we're thinking that it goes from 40 million gigabytes per month available in Gen 1 at the end of 2020 and 25 gigabytes per month of Gen 2. And then that'll go to into four of next year, it'll go to 30 million gigabytes of Gen 1 and 55 million gigabytes of Gen 2 per month can be generated.
Again, capacity is not shipments. We think Micron's underloading the factory, but this is the capacity, so it tells you whether or not we see an issue with being able to get capacity going forward and, right now, we don't. 3D XPoint applications -- you have the fastest NVMe SSDs available on benchmarks, I say, probably because I'm sure people would argue about it, but I think they are. Optane memory for PCs, it's a cash for HDDs and speed storage, millions of units have been sold. Intel announced some QLC NAND Optane combinations and one M.2, cost effective performance versus TLC SSDs.
11:47 MW: Then we have the DIMMs, Intel Optane DC persistent memory DIMMs. In memory mode, you have a large main memory cached in DRAM, but it's not persistent, and originally, some people speculated that this was going to be the main application. It might have started off that way, but the main applications -- the second one which was so-called App Direct Mode -- that's what we call true persistent memory. This is the main focus, and if Intel is going to make breakthroughs on architecture, and when we talk about what should be used in the data center, this is what you want to use it for. It can also act like an SSD on the DRAM bus, so you can use it for block modes, file modes, etcetera. Coming soon, you'll see Optane persistent memory on a new bus CXL that'll be coming out, probably being developed and sampled early on, and then over the next two to three years, you'll see that in production.
12:39 MW: Another question that came up, is 3D XPoint now making money? And the short answer is no. Operating margins are very negative. SSDs were the first product and the sales were all in the Intel NSG group. A number of people did reasonable math that showed losses on XPoint of greater than $300 million per quarter. It was low-volume, expensive development, difficult product, it's really not good for that. NSG losses dropped, but some of this is due to moving the focus, the expenses and the bits to DCG. And also much better NAND prices on enterprise SSDS.
Our model still shows a large negative margin impact to DCG. As persistent memory sales increase, you'll get more and more of that cost and margin impact will go to DCG. On the other hand, none of these matter because persistent memory is great for the data center and Intel architecture and differentiation, and I said from NVMe, but actually from all competition, the margins on storage and memory is not the goal for Intel, it's really selling the platform and all the technologies around the data center. So, the competition for 3D XPoint competing PCM technologies, all companies that worked on PCM, many showed information at recent conferences, we still expect XPoint PCM announcements from multiple companies like SK Hynix, Samsung, etcetera.
14:07 MW: MRAM is great for speed and embedding but can never compete on cost or density. Its competition really is SRAM, not really DRAM or NAND. Resistive RAM, cross-point resistive RAM technologies are the most direct comparison to 3D XPoint cell size, cost, speed, cycling and applications are potentially similar, but the issue is at 2019 IBM and a lack of . . . We saw a lot of issues presented that kind of questioned how fast we're going to be able to ramp resistive RAM. There were a number of new failure modes and we weren't sure about how to fix them, and then there's the lack of new products. We haven't seen any new products on resistive RAM, especially a cross-point array, so we're really no closer to production that we were at FMS in 2019.
Another competition is from fast NAND, so-called low-latency NAND, NAND plus DRAM solutions from NAND and NVDIMM companies will compete in both DIMM and SSD formats and also on a new memory bus, so those are the competition that you should be tracking and making decision on what to choose. Cost and price for 3D XPoint DIMMs, key 3D XPoint features low cost compared to DRAM, this enables more total memory.
15:25 MW: Our estimates to cost were going from a Gen 1.6X DRAM cost in 2022, we think that would go to about 0.45X DRAM cost, and then the cost on Gen 2 will go to about 0.4X the DRAM cost. The prices you see typically on our list are actually DRAM prices or actually they were even higher, if you look at channel prices, and the ASP gave the incredibly vague comment of, "It's less." After looking at some Intel financials, the ASP projection is very difficult. Intel is looking to drive growth in expense of margins, it's also expected that Intel is providing strong incentives for large customers to implement. This can also impact revenue projections, so we know that when Intel wants to implement new technology they'll give strong financial incentives for people to implement it, and as a result, that can affect the price. And I would expect to see wide-ranging prices, depending on whether it's sold in the channel or what Intel is contracting to some of its best customers. So, if you look at 3D XPoint revenue back in August 2019, I said the total revenue would go from $1.35 billion in 2020 to $3.6 billion in 2024.
16:47 MW: I broke it out by DIMMS and non-DIMMS. I updated it this month based on a number of data points, including in some of the finances that came out from Intel's most recent earnings report, and the updated numbers in total are $1.1 billion in 2020, growing still to $3.6 billion in 2024. And one of the big changes, if you go through the details of this graph, I mean, of this chart, is the fact that up till 2022, I think there will be continued limited adoption after 2022 when we move persistent memory to a CXL or other new bus, I think you're going to see a big jump in sales that come after that, and that'll be very enabling across the board, and that's going to be what's going to get us to the 2024 levels of revenue, so it's not . . . Somebody had speculated, I think, $8 billion in 2024 a long time ago, but $3.6 billion is definitely something that Intel's planning for and it's on track.
17:49 MW: This actually includes both Micron and Intel. Micron's expected to be minimal impact until 2022 plus, and one of the other question that came up is, "Where did you get these numbers from?" Some people were asking the same. These are all basically sourced from raw data that I created based on the sales of Cascade Lake and Cooper Lake, what their market share is, what their DIMM attach rate is, what their sale price is for that, and the average opting density. We may get lucky in with the breakout of the SK Hynix NAND sales, Intel may report out revenue numbers that allow us to have more insight into what their actual revenue is.
So, in summary, 3D XPoint ramp has been slower than we projected, but it has surpassed all of the new memory technologies and shipments, Rev-2 3D XPoint is shipping today, teardowns are coming soon. Focuses on 3D XPoint DIMMs for persistent memory. Again, there are SSDs but DIMMs are the focus, that's the breakthrough technology. DIMMs are shipping today, bits are growing steadily since mid-2019, 3D XPoint SSD revenue is expected to grow in the future based on Intel and Micron shipments and new PCIe for products. DIMM revenues is growing steadily and persistent memory revenue will take off at a faster rate with the launch of a new bus architecture in 2022. Total revenue is just over $1 billion in 2020, growing to $3.6 billion in 2024. And that's all I have, thanks.