QLC flash drives are hot '21 trend, pressure HDDs on price

SSDs cheaper than HDDs?

Alex McMullan

Alex McMullan, international CTO, Pure Storage: Disk is going to be more expensive than flash within the next year. I'm not talking about buying a hard drive at Best Buy. I'm talking about what it costs to store data in a data center on a price per protected gigabyte basis, with the RAID overhead.

David Floyer

David Floyer, CTO, Wikibon: The cost of QLC flash drives will be the same as hard disk drives by 2026. The price differential is seven to one at the moment. But the number of HDDs has been dropping dramatically, and because of the lower volumes, the percentage of price improvement is projected to go down each year. As your rate of improvement goes down, the cost of introducing new technology has to be spread over so many years that it's going to cost a fortune. Every year, [heat-assisted magnetic recording] is meant to come out, and it's delayed. They have some [HAMR] drives out there, only in the cloud. But they can't afford to sell them in volume. The componentry is incredibly expensive. The HAMR drives are 20 TB, and they're already up to 18 terabytes using [conventional perpendicular magnetic recording] PMR. So, there's 10% improvement for a 20%, 30%, 40% increase in price. That's not going to work. When you get to 2030, all the HDD output will be to cloud-based environments. Flash will pretty well be the dominant force by far. Tape is going to barrel away at the bottom for sequential work. There's a lot of stuff in the cloud that is just sequential, and tape is much, much cheaper. Exabyte shipments of tape will be twice HDD in 2030. There'll be front-end flash with tape.

Don Jeanette

Don Jeanette, vice president, Trendfocus: The COVID-19 vaccine will come to the masses and, we hope, usher in a sense of normalcy and healthier buying behaviors. Increasing demand from major data center customers -- Microsoft, Amazon, Facebook and Alibaba, etc. -- and PC buyers will bring NAND out of an oversupply situation by midyear and raise prices for flash and solid-state storage. As SSD/NAND vendors charge more to system OEMs, the increased cost could potentially pass through to end users as well, and they could pay more for server and storage rack systems.

Jim Handy

Jim Handy, general director and semiconductor analyst, Objective Analysis: This year is well poised for semiconductor memory prices to stay flat, but there are other ways it could go. Another is that we have a global financial collapse, because of COVID, and prices collapse. DRAM prices could fall to about half of where they are today. For NAND flash, it would be a much smaller drop of about 10%. Because 80% or more of the cost of an SSD is the flash memory inside, SSD prices would likely collapse as well. I'm a doomsday guy, so I'm looking for a pandemic-driven collapse. I look at the fact that we've got all of these unemployed people. We've got a lot of service sector jobs -- everything from airlines to restaurants to movies -- doing immensely badly. That's causing tax revenues to go down, while the government is issuing bailouts. This really doesn't sound like something sustainable. But I'm an engineer by training, not an economist, so take my apocalypse concerns with a grain of salt. 

Enterprises should probably only buy things when necessary, because it doesn't look like there's going to be a NAND flash shortage this year. Demand will increase, but the manufacturers are trying to pace their production output to match it. So, they're still producing more than the market requires. The only thing that could cause prices to go up is if people suddenly stopped worrying about COVID and started buying consumer goods with NAND flash in them. Everybody believes there is pent-up demand, and so do I. But when are people going to turn the faucet back on with their buying of cell phones and other consumer goods that use flash? That's probably not going to happen until the pandemic is fading.

QLC flash coming on strong

Andy Walls

Andy Walls, IBM fellow, CTO and Chief Architect, IBM Flash Systems: You're going to see more QLC -- the four bits per cell NAND flash -- come out in the enterprise and be the NAND flash of choice. QLC is a complicated and difficult technology. It increases the programming time. The latency is longer for reads. The inherent endurance is less. So, all of those things on the surface would leave one to conclude that it simply cannot do as well as TLC and won't be used in all workloads. But there are ways around that. As with most technologies, you put clever engineers and scientists in a room, tell them they can't do something, and they figure it out. If one can cut the cost of the flash by 20% to 25%, it gives options to lower the price to keep a profit margin in a very competitive market. It also ends up reducing the price to the customer either through direct pricing or more capacity inside a single storage unit, rather than having to buy two or three.

Randy Kerns

Randy Kerns, senior strategist and analyst, Evaluator Group: There will be more QLC product offerings this year to the point where almost every storage system vendor has one. The number of NAND flash layers has dramatically increased, and the fab facilities that deliver the chips are coming online now and producing in volume. Once they're producing in volume, we're going to see a significant price reduction and more product availability. This will spur a lot more people to say, 'It makes a lot of sense if I transition away from mechanical devices.'

Sudhir Srinivasan

Sudhir Srinivasan, SVP and CTO, storage division, Dell Technologies: QLC will be truly enterprise-grade in 2022. Yes, QLC is available today. Some vendors are selling it. But we don't think it's enterprise-ready at a cost point that makes ROI sense versus the latest TLC drives. There's not enough economic benefit to outweigh the downsides of the variability in performance and endurance in the QLC NAND media.

Marc Staimer

Marc Staimer, founder and president, Dragon Slayer Consulting: You're going to see more tiered hybrid architectures where you have very high-performance storage class memory in front of low-endurance but high-capacity QLC flash drives. You might even see some customers playing with PLC -- penta-level cell, five bits per cell -- by the end of the year.

McMullan, Pure Storage: Everybody's discussing PLC flash, but QLC and TLC will be the preferred medium for the next three years. If we look at what's happening in the NAND industry, we can see the direction of travel is toward not just more layers but stacking the layers as well, rather than taking that next level of engineering challenge that comes with PLC. PLC has too much complexity. There are too many downsides around endurance and maintaining quality. We're now sort of at an asymptotic level of benefits with looking beyond QLC.

Trends in flash drives and systems

Srinivasan, Dell Technologies: We won't see the first truly all-flash data center until the next year or two. People have predicted we would be there by now, and we aren't. There's a long tail for spinning media.

Eric Burgener

Eric Burgener, research vice president, IDC: We'll see an explosion of enterprise-class systems using QLC media, dropping the price per GB of flash significantly, and we'll start to see strong penetration of all-flash arrays into the secondary storage arena in 2021.

Jeanette, Trendfocus: New PCIe SSD form factors will start to ramp in 2021. Enterprise and Data Center SSD Form Factor (EDSFF) 1U Short and later 1U Long -- or E1.S and E1.L for short -- will start to see growth and eclipse other popular form factors, most notably, M.2 PCIe in the data center market. Certain data center companies will move to the newer form factors for ease of integration, performance, power savings, capacity, etc. As we have seen in the past, if any major cloud providers or hyperscalers migrate to something new, this will move the needle for the form factor since they dictate so much volume.

What's on the horizon for 3D XPoint, Optane

Handy, Objective Analysis: We'll see big growth in Optane because Intel is finally cranking out a broad range of processors to take advantage of 3D XPoint. There has always been big potential for this technology, but a couple of important things have not been around. The most important has been processor support for Optane storage in the DIMM [dual in-line memory module] format. It took hardware changes at Intel to do DIMMs, and so Intel was shipping it in an SSD format. The SSD interface really hobbled the speed of the technology. Now that Intel is shipping multiple processors that can take advantage of 3D XPoint in a DIMM format, that really opens the door for it to be used both as a large memory and also as storage close to the processor. The storage close to the processor takes one other ingredient that is pretty sorely lacking right now, and that's software that takes advantage of it. But, even without that software, for people who are using conventional in-memory databases, Optane gives them a much larger memory than they'd be able to get any other way. So, that's something that will cause wide adoption. DIMM shipments of Optane in 2020 were small enough to be unnoticeable, so revenue could go from near zero to surpass a billion dollars in 2021 if Intel builds the business the way that it should be able to do it. Then it should really kick up a lot from there. SSD revenues for Optane last year were around $400 million, and it's probably going to stay somewhere around where it is or grow very slowly.

Srinivasan, Dell Technologies: Storage class memory -- 3D XPoint, Optane -- is probably at least two years from making sense economically. The performance is great, but the price is so high that it doesn't make sense for most applications. Very few applications need that level of performance. We will probably see greater adoption of DIMMs than SSDs because more and more applications are being developed for memory semantics as opposed to storage semantics. To truly take advantage of persistent memory, the applications have to change to use memory access. Some cloud-native apps are already built that way. The ones that are focused on real-time analytics, for example, assume things are in memory because of the extreme performance and latency needs.

Burgener, IDC: It will be another year of slow growth for storage class memory shipments, although I expect to see most of the established storage array vendors supporting it as an option by the end of 2021. Most of what is sold in 2021 will be sold as devices to go into servers rather than as add-ons to enterprise arrays. We'll start to see a ramp in 2022 as more array customers want it, and because we'll start to see some revenue traction from 'big memory computing' vendors that are better enabling persistent memory operations.

NVMe over Fabrics becomes mainstream

Handy, Objective Analysis: NVMe over Fabric is going to become mainstream. I view it as being something that could be as big of a mover as when virtualization came into being, taking resources that aren't shared and turning them into shared resources in a way that will allow data centers to process data more efficiently at a lower cost.

Srinivasan, Dell Technologies: This is the year where the jury will finally settle the NVMe fabric choice. We'll see ecosystems start to rally around NVMe over TCP. RDMA over converged Ethernet (RoCE) still requires a dramatic amount of network expertise to operate reliably at large-scale fabric settings. The operational efficiency of TCP/IP is going to drive it, and the suppliers will make it work efficiently in terms of performance. The industry has been working towards narrowing the gap between RoCE performance and NVMe over TCP performance, and it's pretty close already.

Thomas Isakovich

Thomas Isakovich, CEO and founder, Nimbus Data: There will be rapid adoption of NVMe-oF using TCP or RDMA over ultra-high-speed 200 Gbps and 400 Gbps Ethernet in a meaningful way. Data-center-grade Ethernet switch costs will fall to under $1.50 per Gb of bandwidth, which is 85% less expensive then Fibre Channel switch costs of about $10 per Gb. NVMe-oF, unlike iSCSI, puts Ethernet storage on par, or faster, than Fibre Channel for latency-sensitive applications, while absolutely annihilating Fibre Channel in performance by six to eight times when it comes to AI, high-performance computing, digital content and analytics workloads that demand bandwidth.

Burgener, IDC: Starting in 2021, many customers will be experimenting with NVMe/TCP in preparation for a ramp of production deployments starting in early 2022. NVMe/TCP will lower the latency of storage access by almost an order of magnitude for iSCSI-attached arrays using NVMe storage. It's inexpensive, using all industry-standard hardware and software on the initiator side. But you will obviously have to have an array that is capable of using it.