Speed is 2020 theme in NVMe flash storage, persistent memory

Flash storage and NVMe over Fabrics

Steve McDowell

Steve McDowell, senior analyst, storage and data center technologies, Moor Insights & Strategy: 2020 will not be the year of NVMe over Fabrics. While nearly every flash storage vendor is delivering some flavor of NVMe over Fabrics, we won't see mainstream adoption until at least 2021. The technology is too new, too complicated, and will be driven by IT storage replacement cycles more than greenfield deployments. This is a long-term play, not one that will quickly resolve itself in 2020.

Eric Burgener

Eric Burgener, research VP, enterprise infrastructure practice, IDC: Between 2021 and 2023, we will see more NVMe-oF implemented using the TCP transport protocol rather than other Ethernet options like RoCE and iWARP. The maturing of the NVMe over TCP standard, as well as products based on it, will help drive the growth, particularly for greenfield deployments where Fibre Channel storage networks are not already installed. What makes NVMe over TCP so interesting is that it will be entirely based on industry-standard hardware and software that ships with most commodity Linux servers. RoCE and iWARP require custom content to support the [remote direct memory access] RDMA capability. NVMe over TCP will provide most of the performance advantages of the other Ethernet transport options in an industry-standard package that will be cheaper to buy and easier to deploy and maintain. 

Hu Yoshida

Hu Yoshida, VP and global CTO, Hitachi Vantara: Increasing implementation of NVMe on the front and back end of the storage controller will impose bottlenecks in traditional storage controllers that will limit the realization of NVMe's full potential. It will also impact other software functions that enterprise storage controllers provide, such as deduplication and compression. This will require a new generation of storage controllers that are specifically designed for the parallelism and low latency of flash and NVMe and NVMe over Fabrics.

J Michel Metz

J Michel Metz, field CTO, Rockport Networks: The age of new NVMe investment is coming to a close. A couple of years ago, it felt as if all you had to do was say "NVMe" and there was a new startup every week. Now, new ideas are going to be hard to come by, as block storage like NVMe has a finite number of creative applications. Most of the missing pieces have been found, and there will be far fewer new investments in different form factors. Once again, the momentum will shift toward software.

Persistent memory heats up

Jim Handy

Jim Handy, general director and semiconductor analyst, Objective Analysis: 2020 will be an interesting year because we will finally start to see applications that use what SNIA calls persistent memory (PM), and many folks call Storage Class Memory or SCM. This will largely be embodied as Intel's Optane DIMM, which officially bears the unwieldy name of Optane DC Persistent Memory Module. What's important is that this hardware, when coupled with PM-aware software, will create extraordinary speed improvements to certain software, and this could cause a dramatic change to the way that computer memory is configured, just as SSDs redefined storage hierarchies.

Andy Walls

Andy Walls, IBM fellow, CTO and chief architect, IBM Flash Storage: 2020 will be the breakout year when Optane DIMMs and storage class memory finally become more widely used. I think NAND price drops are going to reverse, and we'll start seeing price increases in 2020. Optane and storage class memory prices will look more attractive. I doubt that we'll have many storage systems sold with all SCM, but you'll start seeing large vendors like Dell, IBM and others have storage class memory in a tier that can provide performance improvement. In 2020, it's going to be mainly NVMe SSDs. But Intel's also enabling Optane DIMMs, and you'll see some proofs of concept. Storage services like compression and data reduction create a ton of metadata. If that metadata can find its way to something like SCM, I can reduce my response time considerably, provide more consistent response time, and help my service-level agreements.

Tom Coughlin

Tom Coughlin, president, Coughlin Associates: We're going to see the first significant commercial applications of MRAM technology in embedded devices. All the foundries have said they're going to make it. This year, we're going to see the production of real products used for IoT and artificial intelligence types of applications. MRAM has very high endurance, high performance and low latency. SRAM takes up a lot of space on the semiconductor die. MRAM is smaller, so it gets more memory in a given amount of space. It's also non-volatile memory, versus SRAM or DRAM. That means I can turn the power off, and my data is still there. MRAM allows lower power usage. That's particularly important for some embedded, IoT and other applications that may be running on battery or low power.

Flash storage price trends

Burgener, IDC: By 2021, NVMe-based all-flash arrays will drive over 50% of all primary external storage revenue. As many enterprises move to data-driven business models, they will often be implementing new workloads that require more performance than SCSI-based AFAs can deliver. Many established storage vendors anticipated this and introduced NVMe-based versions of their flagship enterprise storage arrays that are priced on par or below the prices they are charging for their older, SCSI-based arrays.

Handy, Objective Analysis: DRAM and NAND flash memory chips continue to be oversupplied, but while NAND flash hit bottom in the middle of last year and is now selling at cost, DRAM is still selling at about a 30% gross margin. Since DRAM is oversupplied, that margin should go to zero in the first half. This is a boon for computing applications that need larger memories or more storage. NAND-based SSD prices are close to where they will remain for the rest of 2020, but dual in-line memory module (DIMM) prices will continue to fall until they reach cost, which is below $2 per gigabyte. These two technologies will not see a big increase in consumption in response to their improved availability and low prices. That almost never happens. Instead, demand will continue to grow along its past trajectory, and price benefits are likely to get passed down to the end user.

Coughlin, Coughlin Associates: We're going to see 96-layer 3D NAND flash in volume production, and we'll see an introduction of 128-layer flash and probably the announcement of 192-layer flash. I don't think 128-layer flash will reach volume until 2021. It's higher density and therefore lower cost per bit. People should see cheaper flash, but it all depends on volume and availability.

SSD developments

Phil Bullinger, SVP and GM, Data Center, Western Digital: Storage is becoming more workload-optimized, especially for sequentially written, read-intensive workloads such as video or IoT streams. Up until this point, applications and operating systems have treated storage devices as simply random-write, random-read. Zoned Storage is an emerging initiative where the applications and the operating systems have more intelligence and communicate with the I/O device, writing data sequentially into blocks of storage.

By optimizing, you can make the storage more efficient, higher performance and more effective. In 2020, Zoned Storage will come to fruition -- between the standards activity to build it into the NVMe protocol, multiple SSD suppliers with samples and development-level product and pull from large cloud-scale customers and branded server players that also have storage components. Zoned Storage is similar to shingled magnetic recording (SMR) in hard disk drives. SMR technology is built for exactly the same purpose: sequentially written, read-centric workloads.

Burgener, IDC: The use of all-flash arrays (AFAs) will begin to take off by the end of 2020 for secondary storage workloads including test/dev, backup, disaster recovery and active archive. While low latency is not generally required for these types of workloads, solid-state storage characteristics like high bandwidth/throughput, higher storage density, faster execution of data services, and higher system reliability can still be very attractive. Wider use of dense new solid-state media like quad-level cell NAND flash will enable this, as it narrows the price-per-GB differential between flash and 7200 RPM HDDs. But it will require new software algorithms to ensure the QLC flash media endurance meets enterprise requirements. Early on, those vendors that buy NAND flash media off the shelf and design their own devices will be able to meet those enterprise requirements more quickly and for a broader set of workloads than off-the-shelf SSD vendors.

Don Jeanette

Don Jeanette, vice president, Trendfocus: QLC deployments in SSDs will remain limited in 2020, with only a couple vendors pushing this aggressively. QLC production is still limited. Also, customers don't like to transition in volume to another technology when performance takes a hit. Performance of QLC-based SSDs is much lower than other NAND technologies. 3D TLC will remain the significant volume NAND technology throughout 2020 and into 2021. It remains a cost advantage over competing technologies, and it's a known entity that is already qualified and shipping in most SSDs in the world.

M.2 and U.2 form factors will have the highest PCIe volume this year. Dual-port PCIe remains limited in the storage networking world. This means SAS SSDs will continue to have a healthy life in the near to mid-term. Vendor roadmaps for SAS SSDs remain robust. Also, the higher performing dual-port PCIe solutions will remain at a higher price tier than SAS-based systems. When you are the higher priced tier, volumes will remain lower.