Download the presentation: NVMe Market Research
00:01 Mike Heumann: Hi, welcome to the Flash Memory Summit Virtual 2020. My name is Mike Heumann. I'm the managing partner and principal storage analyst for G2M Research and the topic we're going to be talking about today is NVMe and NVMe over Fabrics, progress over the last five years and preparing for the next five years.
So, a little bit about G2M and G2M Research. Last year, if you remember at Flash Memory Summit 2019, we actually did the "In the Hotseat" interviews with about 13 companies. Obviously, this year -- a little bit different situation -- not going to be able to do that this year, but perhaps we might do those early in next year, we'll see. Certainly, if the Flash Memory Summit goes physical next year, we'll be back and we'll be doing them there, too.
So, G2M Research is really an outsourced marketing firm -- we do marketing consulting and sales consulting for companies. One of the things that we do as a part of that is market research, so we've always focused on emerging markets, NVMe and NVMe over Fabrics is one of those markets, it's part of the larger enterprise storage market which we focus on.
01:15 MH: We also do work in areas like cybersecurity and artificial intelligence, in artificial augmented reality and virtual reality, and then electronic vehicle charging infrastructure. Those are just some of the areas we cover, there's other areas as well. And from a market research standpoint, we don't try to do the typical, "Hey, ask the vendors how many they sold of X, Y or Z." What we typically try to do, rather is look at the whole ecosystem in general and say, "OK, how do all these numbers fit together?" Because if you do that, then you can understand a little bit better what's driving a market, and that certainly helps our customers understand where the opportunities are. So, that's kind of the approach we take. And if you want to find more out about us, you can contact myself, [email protected].
This article is part of
02:10 MH: So, we had several predictions that we made in 2016, which was the first year we did our NVMe market sizing report. That report actually debuted at Flash Memory Summit 2016, and we had six interesting predictions. Let's see what they are and how they did versus actual reality.
So, the first prediction we had is that the NVMe market -- and when we say market, we mean everything from servers that are NVMe capable, to storage arrays that are NVMe capable, to storage adapters or networking adapters to SSDs, to software, so the entire ecosystem -- we saw would be $57 billion by 2020. The second prediction we made was that over 50% of servers will ship with NVMe drives or NVMe slots in them by 2020. The third prediction: Over 60% of storage servers will have those drives and slots in them, they will be NVMe slots by 2020. So, a storage server being a server with a large number of slots for storage devices -- think of 16 or 32 slot boxes, 64 box slots that you'll get from like a Supermicro or people like that, or other OEMs or ODMs.
03:32 MH: The fourth one is that NVMe over Fabrics adapter shipments will exceed 740,000 units by 2020. Fifth was, all-flash arrays will move to NVMe and 40% of those arrays will be NVMe-based by 2020. And the last one was that NVMe SSDs will reach pricing parity with SATA SSDs by 2018.
So, with that, let's take a look and see how we did on those predictions. The first one, if you remember, was the size of the market. So, our current estimates put the NVMe market at around $60 million, by given COVID and everything, that's actually pretty good. So, if anything, we underestimated the market. Secondarily, is that nearly all servers shipping today support M.2 drives. So, if you remember we predicted that 50% of servers will ship with NVMe drives or NVMe slots. Basically, today 100% of them do, if you look at M.2 alone, if you look at U.2, it's closer to the 50% number but, so today obviously, overwhelmingly most of the servers that ship, support NVMe.
04:44 MH: The second prediction had been over 60% of storage servers will support NVMe by 2020. That one it's a little harder to say. They don't generally release numbers, but if you look at the storage servers that are out there, certainly by a model count, they're predominantly NVMe and certainly given the fact that more than 50% of all enterprise SSDs that are shipping today are NVMe, it's pretty safe to assume that it's likely that over 60% of the storage servers would be NVMe as well.
On the adapter side, if you look today at adapters and network adapters -- whether they're InfiniBand, whether they're Fibre Channel, or whether they're Ethernet -- nearly every single adapter supports NVMe over Fabrics. That doesn't mean they're all used that way by end users, but it means the capability to do NVMe over Fabrics exists in every single one of those adapters today. And the next one was, over 40% of all flash arrays will be NVMe-based by 2020. If you look at the market that's out there, certainly over half of all the all-flash arrays utilize NVMe drives at least in part and support NVMe over Fabrics networking.
05:57 MH: Again, it doesn't mean that they're used that way by end users, but that is what is out there in the industry and that's what's available to end users. And then, finally, on the pricing parity of NVMe and SATA SSDs, the SATA market companies have driven the pricing down significantly on their drives, so we're not quite at price parity. But if you look at how much NVMe SSD prices have decreased, you can see that at a point they almost match what the SATA SSDs were at, they're certainly within 10% to 15% depending on model, type, performances and other factors. Invariably, you can tell that they certainly reach a point where the price is certainly reasonable for the average user, because today NVMe drives, in fact, that crossover happened in 2018, NVMe drives constitute the majority of SSDs that are sold into the enterprise.
06:56 MH: So, let's look at a couple of reasons why NVMe took off so quickly. It certainly got a lot of user adoption in a very quick fashion. One of the first ones is performance. If you look at NVMe drives, the performance on these drives, especially in the case of latency, is significantly greater than the performance of similar SAS, SATA SSDs, and a lot of that has to do with the fact of SAS, SATA SSDs, you have a very large protocol stack that you have to process before you can move in. With NVMe, it's much simpler; now that makes it much faster. Secondarily, from a capability standpoint, NVMe provides a lot of features not available in SAS, SATA. Think about things like zone namespaces, virtual SSDs -- there's just a whole bunch of features that it has that really are built around flash, that SAS and SATA just don't support because they're built around rotating media.
08:00 MH: And for a lot of these features provide a lot of capabilities to people that are using these as systems for storage purposes or using them for application acceleration, and that's very important for those users. Finally, pricing. As we said a little earlier, the pricing on NVMe SSDs is close enough to SAS, SATA pricing. That's induced switch-over. Certainly, from a price value standpoint, you can say NVMe SSDs have provided more value for the price, and that's why if you look at some other market spaces like the gaming space, laptops, things like that, NVMe SSDs are almost completely all you see. They are, by far and away, the lion's share of the market in those specific markets.
08:49 MH: One of the things that's big about NVMe and NVMe over Fabrics is they've really enabled a lot of breakthrough concepts the companies have been able to use to improve both performance and some other capabilities of their applications and of their architectures, the storage architectures. Well, a good example is what we call scale of flash storage software, or SOFS software. You can see some of the names and some of the companies up here that do this particular type of software. And the good thing about SOFS software is it allows you to take a bunch of servers, each have their own storage pools built out of NVMe SSDs, and if you have NVMe over Fabrics networking connecting these, you can combine all the storage pools into one large namespace.
09:40 MH: That namespace then you can partition up while certain uses these certain parts. All the type of capabilities that you've always had with storage arrays you have with this type of software. It's really kind of software to find Storage 2.0, a lot more capabilities than the old software finds storage opportunities had, and certainly has seen a lot of pick up by a lot of large enterprises.
The second one here, virtual SSDs. So, that allows you to take a solid-state drive that's an NVMe drive and split it up into multiple virtual SSDs that you can assign different sizes to, different shapes to, you can actually give them different quality of service settings, and you can vary all this over time based on software.
10:31 MH: That is really valuable especially when you think about the fact that these SSDs keep getting bigger and bigger. Now, we have SSDs now with tens of terabytes -- 16 terabytes isn't uncommon -- we're going to have 32 terabyte SSDs. And for a lot of cases, splitting these up between workloads helps manage them much better.
The third one is composable storage architecture. So, if you think about previously, say 10 years ago, if I wanted to build a storage array that was disaggregated, what I typically had was a couple of controllers sitting at the top of a rack and then they connected with typically the SAS to a bunch of JBODs or JBOFs that were underneath them.
11:14 MH: OK, while this worked and allowed you to extend your storage pools, it didn't provide a lot of flexibility because those connections were one-to-many connections, it wasn't a true network. Now, with NVMe over Fabrics, you can actually have a true network on the back of your storage. If you look over here, this is an example of a different setup. You can see this has three controllers and they're all sharing five different JBOFs, and you can move these connections around because this is all one big fabric. If it's done with InfiniBand or Ethernet or Fibre Channel, you can connect any given JBOF to any different controller and you can move them around as you need. That increases your performance, but it also increases your security, so there's a lot of value to that as well.
And then finally, the enterprise data center small form factor SSD, which has recently been introduced over the past couple of years, and you can see here an example of one by Intel.
12:14 MH: These are also built by Kioxia, Samsung, and all the other guys that are building SSDs. This is a form factor that's been built specifically around flash, unlike our 2.5-inch U.2 drives, which are nice. They provide a lot of capability, but those were built around 2.5-inch rotating disks, and they didn't necessarily optimize the space of a footprint for flash cooling or for flash, just total density. EDSFF does that, and it does it in such a way that really increases the capacity. It also -- there, if you look at this particular one, it's a 1.0 form factor, this is optimized for one-use servers. Now, this allows you to put 16 drives or more into a one-use server. You can't do that with U.2 drives, so there's a lot of value to that as well.
So, we kind of covered some of the different things that NVMe has done. So, what does that mean if you're looking at a data center, you're a data center storage architect or a storage network architect, what can you expect of NVMe or NVMe over Fabrics in the next five years?
13:23 MH: Now, the first one is SSDs. Expect these SSDs to move from U.2 to M.2 to EDSFF over the next five years. We actually think that'll start getting a lot of steam in the next year to two years. And we expect by 2025 that 50% of the enterprise drives that we sold will be EDSFF drives.
Here, you can see an example of a 1U EDSFF-enabled storage server -- it has 24 drives in it. If you think of each of these drives as being 8 terabytes or 16 terabytes, that's a lot of storage. And with the way the storage densities are going, it's foreseeable that you could have a petabyte in a 1U storage server by 2025. Secondly, NVMe over Fabrics will be the dominant storage networking technology by 2025.
14:17 MH: As we said, every single adapter out there is NVMe over Fabrics capable. What we don't have yet today is the majority of the users actually using NVMe over Fabrics because most of these adapters support more than one protocol. They can support NVMe over Fabrics-based storage protocols, they can support iSCSI in many cases, and if they're on the Ethernet side, obviously the Fibre Channel ones could support old Fibre Channel protocols or they can support NVMe over Fibre Channel, same with the InfiniBand drives. So, it's really a matter of the different applications and different end users switching over and utilizing NVMe over Fabrics. That's really the change that we need to see over the next five years to hit this particular milestone.
15:04 MH: And then No. 3 is that the scale of flash storage software -- and I'll show it here kind of underlying NVMe over Fabrics since it actually runs on top of it -- but scale-out flash storage software using NVMe over Fabrics will approach or exceed the amount of storage capacity that is sold in the market, that is coupled AFAs today.
So, if you think about in five years, more storage will be connected by scale-out flash software than will be connected to end users through all-flash arrays. And that's simply because you see a better price value point and performance-wise, it has a lot of the same characteristics, where in some cases better characteristics than all-flash arrays have. So, if you're an enterprise data center storage architect, you might be asking yourself, "Hey, this is great. How do I prepare for this? What can I do to make sure I can take advantage of all these capabilities?"
16:02 MH: So, let's talk about the first one, and that is buy the largest drives that meet your performance needs. So, if you've got a specific performance profile you need, buy the biggest drive you can get on that performance profile, and you may not need the space today, but in two or three years, you will and you don't want to have to go out there and retrofit those servers with larger drives or return them ruined because you don't have enough storage, that would be a bad thing.
Secondly, over the next couple of years think really hard about making that transition from U.2 drives and M.2s to EDSFF form factors. Do this even if your server vendor doesn't tell you that you should do it. Ask them about it and say, "Hey, what about this? What do you think? Is it time to start moving over to this?" Because they're going to come and the performance is going to be a lot better. Density is going to be a lot better, and it's a good way to scale your data center capabilities and also reduce your costs and your footprint.
17:04 MH: And then finally, you know, one of the things we spent a lot of time on over the past 10 or 20 years is deduplication or eliminating duplicate data within the data center. When storage was expensive and hard to come by and performance wasn't so great, that was a big issue. Today, you've got huge drives that can store lots of data. So, there's no reason to be afraid of that. Networking speeds, even for storage networks, are such that you can move that data around very easily and if you think about a lot of your workload being virtualized and things like that, having duplicate data sets probably makes both virtual machine migrations and, you know, things like data availability and disaster recovery far simpler.
17:53 MH: Also, by having duplicate sets of data, it allows you to have different sets of data along with different applications, and that means you can reduce the amount of hotspots you have in your network, which is also very valuable. And it also simplifies data management. You know, you might think that hey, the more copies of a piece of data I have, the harder it is, that's actually kind of a converse because you have to worry about it as much. As long as you keep that data synchronized, it's actually not a bad thing. And then a couple of final thoughts out here, that I can just throw out there, we talk a lot about in our technology world about disruptive technologies. In reality, most technologies that are successful -- probably nearly all of them -- aren't disruptive, they're evolutionary. If they're disruptive, they'd be really hard for end users to actually pick up.
18:44 MH: So, you know, if you look at NVMe and NVMe over Fabrics, they're not by their own nature disruptive, they're evolutionary. It's applying new networking technologies and new interface technologies to products that we already have. On the other hand, though, they do enable disruptive changes in applications and in your data center. So, as you start thinking about these things and you look at how you can use this, be flexible and think outside the box.
There's an old saying that sharks keep swimming or they sink or they drown, right? You don't want to be one of those sharks at the bottom of the ocean -- you want to keep moving and staying ahead of your competition. And NVMe and NVMe over Fabrics can be a great way to do that.
19:26 MH: Secondly, as storage capacities grow exponentially, think about what you can do with that data, with all that space that you got. We talked about duplicate data sets as one option for improving performance and availability, there's other things, obviously data mining, business intelligence. As the internet of things explodes, as you gather more data on your customers, it's very likely that you're going to have a lot of data. Don't throw it away; store it -- you might need it in the future and having it's a lot better than not having it. It's interesting because Tesla recently said that one of the biggest advantages they have over their competitors in the car market is the fact that they have more data than anybody else.
20:07 MH: So, that applies not just to the automotive industry, but to every industry in general. And then, finally, stay informed. You know, there's a lot of different ways you can do this. G2M Research has a monthly webinar on different subjects. We invite multiple panelists in to talk about these things on each of these different webinars -- it's a great way to keep informed. Other organizations also do webinars, too, so there's a lot of content out there.
You might think in the days of COVID that it's harder, in virtual trade shows, that it's harder to actually go get that content, but it's actually easier than you think. There's a lot out there, just to look for it. If you want more info on what we do on some of our webinars or research papers, you can find it at www.g2minc.com/research.
And with that, I'd like to thank you very much for this. It's been great being a part of the Flash Memory Summit Virtual 2020, and we look forward to a Flash Memory Summit in-person next year. Thank you very much.