The year ahead promises to be another fascinating time for the storage industry. In this first blog looking at predictions for 2021, we post some quick thoughts on the media part of the business, both from a micro and macro view.
Plus Ça Change
At a high level, the storage industry continues to do what it has done for decades. Device capacities increase, while prices decline on a $/GB ratio. Hard drives are pushing into the 20TB+ domain, with much larger drives continuing to be developed. Solid-state disks grow in capacity with the adoption of increased layer density and the prospect of penta-level cell technology (PLC). Tape media continues to increase in capacity and incrementally in performance. In 2020, Intel announced second-generation Optane, with increased capacities and modest performance improvements.
The effect of incremental media improvements is to generate a multi-tiered industry. At the high end, Optane in SSD format is starting to replace SSDs as the performance tier of choice. The time for this transition to occur is simply a price/performance/capacity equation. We might include Samsung Z-NAND and Z-SSD in this group too.
As NAND chip density improves, TLC, QLC and eventually PLC hybrid drives will increase device capacity at the expense of performance and I/O density. In the HDD market, the restrictions of mechanical movement resulted in a rapid decline in I/O density for hard drives, with over-provisioning and short stroking the order of the day. NAND flash is heading in the same direction, albeit at a slower pace.
Intel Optane has taken some time to become established. The overly optimistic headline performance figures may have had some bearing on this, but in reality, it’s more likely that processes had to be improved to make Optane technology close to break-even or profitable. We’re starting to see new players including Everspin trying to introduce persistent media, such as MRAM into the market. We can expect these technologies to have a limited impact until the cost equation makes sense.
What can we expect, specifically, in 2021? It’s difficult, if not fool-hardy, to call what specific events will happen this year and even more unpredictable with the effects of the ongoing coronavirus pandemic. However, we can perhaps make some predictions.
PLC Flash emerges. PLC or penta-level cell NAND flash improves the number of bits stored in a single cell from four to five. This evolution represents a 25% improvement over QLC technology and a continuing declining return in percentage terms.
- SLC -> MLC – 100% increase in capacity
- MLC -> TLC – 50% increase
- TLC -> QLC – 33% increase
- QLC -> PLC – 25% increase
Each successive improvement in bit density is countered by a reduction in endurance, meaning PLC technology will best serve read-intensive workloads. Storage vendors have made significant improvements in extending the endurance of media. However over time, the gains in bit density also introduce slightly worse latency than the previous generations of NAND.
I don’t expect to see PLC-only devices arrive just yet. Instead, PLC technology makes more sense to combine with solutions such as zoned storage, where a portion of the media operates in PLC mode and other areas in TLC or QLC modes. This approach may get combined with other solutions like computational storage (more on that in a moment). So, expect to see multi-tier and dynamic SSDs from 2021 onwards.
400-layer NAND flash. Today’s products are pushing 176 and 192 layers with no apparent let-up in the layer count that can be achieved. At a Flash Memory Summit event a few years ago, a Toshiba representative indicated that 500+ layers were practical, so we can expect at least announcements of 384 or 256-layer products in 2021.
If layer density can be increased without significant re-engineering, then the returns are greater than those seen from increasing bit density. At some point, however, the doubling factor may reduce, putting a practical scaling limit on 3D-technology. For now, 3D-NAND will be the primary process used to increase device capacities.
Optane 3.0 and Micron QuantX. Intel recently announced version 2 Optane solutions. At the same time, the Intel NAND business is being sold to SK Hynix. Intel is betting squarely on Optane being the long-term future for persistent storage and persistent memory in the data centre. As a result, the price of Optane needs to drop, and capacities need to increase to the point where Optane challenges high-end flash and is cheaper than DRAM in persistent memory format.
I expect 2021 will see at least an announcement of Optane 3.0, with improved capacity and performance metrics. I want to justify using Optane more in the lab, and I think 2021 will be the year to achieve that.
What about Micron? With increased competition in the processor market, the time is right for Micron to sell their version of 3D-XPoint to a market that includes AMD and Arm. Although the persistent memory version of Optane requires a new memory protocol (DDR-T), Micron must be able to compete in the SSD market and have some plans to integrate with AMD and ARM architectures. It’s the perfect storm. I’d make a Micron/AMD/Arm partnership an outlier for 2021. An alternative opinion is that Intel exits the solid state storage business altogether.
Zoned Media. We’ve already discussed zoned SSDs, which are one way to take advantage of lower-endurance NAND. In both the HDD and SSD market, zoned storage will evolve to be a requirement when taking advantage of large-capacity media. SMR technology allows greater bit-density for hard drives with a trade-off that introduces a sequential write mode. Today’s operating systems are already built to optimise for SMR, but as we saw in 2020, transparency is needed from vendors when introducing technology that has widespread performance implications.
HAMR, HDMR and MAMR technologies will continue to expand the roadmap for increasingly larger HDDs throughout the 2020s. Dual actuator drives have been introduced by Seagate and effectively zone a physical disk into two independent drives. If the costs can work, we may see multi-actuator drives in the next few years, with perhaps one independent read/write head per platter.
- #161 – Seagate MACH.2 Dual Actuator Drive Deep Dive
- #174 – Introduction to Zoned Storage with Phil Bullinger
All of these technologies point to the need for greater parallelism in external peripherals – a trend that is set to continue over the next decade and has been assisted by the introduction of NVMe.
Computational Storage use cases. As a solution to the I/O bottleneck, computational storage offers an exciting approach that pushes compute to the storage media device. In many respects, the solution is effectively distributed computing, where the code executed performs some standalone process on the data stored.
Computational storage vendors already have products on the market, but the application use cases so far, seem limited, with hype greater than practical reality. Early solutions are either implementing batch or event-based processing, or implementing storage functions such as compression in embedded firmware or FPGAs.
For 2021, we need to see a new programming model emerge that allows distributed processing across multiple devices. We also need improved security controls to ensure that computational storage devices can’t be used as an attack vector.
I expect to see computational storage gaining an initial foothold at the edge and integrated into traditional storage appliances as an offload for data-intensive tasks.
DNA Storage. The concept of using DNA for long-term storage has started to gain momentum and column inches. I don’t see this technology going anywhere in 2021 or the next decade, other than to remain a research project.
The Architect’s View
We can expect more of the same in 2021, with few big pieces of storage media news. Not much has changed in 12 months, but looking back over ten years, the storage industry has been revolutionised by adopting solid-state media in all its forms.
Probably the most significant change in the last decade has been in the increasing level of integration and co-operation between storage software and hardware. HDDs and SDDs can’t be treated as a black box that just responds to I/O requests. The physical characteristics of each medium have to be considered.
In many cases, these dependencies are obfuscated by either the O/S or another layer of application indirection. In the coming decade, this complexity is set to increase, with successful solutions and products interacting more closely with the media on which they store their data.
Copyright (c) 2007-2021 – Post #7ca8 – Brookend Ltd, first published on https://www.architecting.it/blog, do not reproduce without permission.