Flash memory cost reductions: getting harder, but still happening
September 9 2019
by Tim Stammers
The annual Flash Memory Summit (FMS) in Silicon Valley in August is a major industry event, and this year it drew about 5,000 attendees. Despite the current glut in flash chip supply and the resulting collapse of prices and reduction in chipmakers' revenue, the major themes at this year's conference included the continuing efforts to reduce the price of flash, as well as boost its performance. The latest step change in costs is being delivered with the entry of QLC (or four-bit-per-cell flash) into datacenters, but it is not yet known whether five-bit or penta-level cell (PLC) flash will follow suit at some point. At the conference, Western Digital (WD) and SK Hynix promoted a zoning mechanism based on open standards that changes the way flash drives interact with host servers and promises to address the poor write performance of QLC, as well as reduce overall drive prices. Toshiba Memory (soon to be renamed Kioxia) and SK Hynix also described developments within flash chips.
Toshiba Memory said its XL-Flash high-performance flash variant will enter volume production this year, while SK Hynix predicted that flash chip capacities will grow eightfold between now and 2030. Toshiba cited a prediction that flash demand will more than quadruple over the next four years, rising to 281 exabytes in 2023. Already, clouds account for over half of flash consumption, and by 2023 they will account for about two-thirds, according to that forecast.
The 451 Take
Whether in the short or long term, forecasting flash supply and demand (and, hence, pricing) is very challenging, as demonstrated by the conditions that chipmakers are suffering at present. However, despite the current downcycle, the rate at which flash prices will fall in the longer term looks likely to be less steep than in the past. One of the major ways of cutting per-GB costs has been to store more data bits in each memory cell, but the returns for doing that are diminishing. The zoning concept WD is promoting appears to be a promising way of extending that process, making QLC – and perhaps future PLC-based flash drives – cheaper and suitable for a wider range of applications than at present. However, it is not entirely clear what level of performance boost will be delivered by zoning, and only two major drive makers are publicly supporting it so far. Another way of reducing per-GB costs has been to increase the number of layers within 3D flash chips, but while this is expected to continue for several more years and generations of chips, it involves ever growing capital spending. The same applies to the reduction of the physical memory cell sizes. On the bright side, China has declared a strong ambition to become a major chipmaker, but is unknown how long it will take to close its current technology gap with existing flash suppliers, or whether the major US market will be off-limits to Chinese chips.
Zoning into faster, longer-life QLC
The recent industry transition from 2D to multi-layer 3D NAND flash chips refreshed the technology curve for flash, and not only by sidestepping the problem that chipmakers had reached the limit in how far they could continue cutting per-GB costs by reducing process geometry sizes and squeezing more memory cells into each chip; 3D also allows each cell to reliably store more data bits, and enabled the move from two-bit MLC to three-bit TLC flash, providing another means of cutting costs. Layer counts are rising quickly, with chips now comprising 96 layers, and the next generation of chips with triple-digit layer counts promises to emerge this year. But as Western Digital said in its keynote, the process of adding layers is already hugely capital-intensive and is becoming more so. And although four-bit-per-cell QLC flash is now entering the market, the cost reductions achieved by packing more data bits into each memory cell are diminishing.
About four years ago, the move from first-generation, one-bit SLC to two-bit MLC flash doubled the number of bits in each memory cell. In contrast, QLC is now increasing the bit count by only 33% compared to TLC flash, while also reducing performance and flash write-life, or drive service-life. Compensating for the latter increases costs elsewhere, for example, by requiring more front-end DRAM cache and more flash overprovisioning within drives. As one way to address those problems, WD is promoting a mechanism called Zoned Namespaces (ZNS), which it says will boost the low write speed of QLC and extend drive lives by using flash more efficiently and reducing write activity. 451 presumes ZNS is a major factor in WD's prediction that by 2025, QLC will account for about half of all flash capacity shipments.
According to WD's keynote at FMS, ZNS is already supported in Linux and is in the process of becoming part of the NVMe standard. This was not WD's first pitch for ZNS or involvement in zoning. In June, WD unveiled an initiative intended to promote zoning in both disk and flash drives. Called Zoned Storage, the initiative is being backed by suppliers that include ATTO Technology, Broadcom, Mellanox and SUSE. The zoning concept has already been used to boost the write performance of shingled magnetic recording disk drives, including those made by WD. Because ZNS-based zoning of flash drives is similar, WD says only limited work will be required to create the ZNS standard and adapt third-party software to support it. ZNS embodies the open channel concept that has already been implemented by hyperscalers, and servers directly manage the capacity of flash drives rather than leaving that to be handled by the flash translation layer built into conventional flash drives.
WD says it is already shipping zoned flash drives to selected customers. During its keynote, the drive giant did not say how much ZNS will boost write performance, but SK Hynix, another major flash supplier, said in its keynote that the mechanism cuts write latency by 25%. The Korean flash chip and drive maker demonstrated zoned flash drives at FMS and predicted that ZNS drives will be deployed by hyperscalers in 2020. However, neither WD's nor SK Hynix's devices will be the first zoned flash drives on the market; memory specialist RADIAN began shipping zoned drives in 2018. RADIAN claims that it introduced the zoning concept to the storage industry in 2015. Samsung, Intel and Micron have made no public statements about ZNS.
Developments inside flash chips
Whether or not penta-level (PLC) flash will see volume production has not yet been decided. It has been publicly demonstrated by suppliers such as Toshiba Memory (soon to be renamed Kioxia), but like others, Toshiba hasnâ€™t committed to putting it into production. Toshiba Memory's keynote highlighted other ways of reducing flash costs, which include a NAND chip architecture under development that it said could effectively halve the size of memory cells and, therefore, significantly reduce per-GB prices. Toshiba Memory also said it has developed a feature based on 'virtual multi-LUNs,' in which multiple planes within a 3D chip operate independently to deliver almost doubled random read performance, according to the chipmaker.
In its keynote, SK Hynix predicted that by 2030, the industry will be producing 4TB NAND flash chips, with eight times the capacity of the current largest chips. Part of this increase will be achieved using periphery-under-circuit (PUC) architectures, which see supporting circuitry moved from outside to underneath a NAND 3D stack. SK Hynix is using the label 4D flash to describe its combination of PUC and charge-trap NAND memory cells, but these technologies are not unique to Hynix. Charge-trap NAND is manufactured by most of the other major flash suppliers, although Intel and Micron were still using floating-gate NAND architecture until the two companies ended their flash-making alliance in 2018 (after which, Micron said it would move to charge-trap NAND.) In 2017, Micron claimed to be the first flash maker to put PUC into production.
Toshiba Memory's faster flash variant
Last year at FMS, Toshiba Memory announced the development of what it has branded as XL-Flash, a NAND variant intended to rival Samsung's Z-NAND flash, and less directly, Intel's Optane memory. Like Z-NAND, XL-Flash is based on SLC NAND, which stores only one bit per memory cell and is, therefore, more expensive than the TLC triple-bit flash that now dominates the market, but is also faster and enjoys a longer write-life.
This year, Toshiba Memory announced that NVMe drives powered by XL-Flash will ship before the end of 2019. Toshiba claimed read latency of just five microseconds for the drives, which is about 10 times faster than typical TLC-powered enterprise flash drives, and about the same as NVMe drives powered by Samsung's Z-NAND and Intel's Optane memory. However, such simplistic claims are only loose indicators of true performance. Like other performance parameters, latency can vary widely in operation and is highly dependent on IO profiles and other factors. As a result, performance comparisons and claims are only valid when workloads have been defined by multiple parameters.
More importantly, Intel is not expecting NVMe drives to be the biggest selling vehicle for its Optane memory. Instead, Intel expects the Optane Data Center Memory Modules (DCPMMs) that it launched earlier this year to sell in significantly larger volumes. The latter are a form of NVDIMM and are far faster than any NVMe drives, with a claimed average read latency of just 0.2 microseconds. That is because they are linked to CPUs via the high-speed DDR interface designed for use with DRAM, and not the slower PCIe-based NVMe interface. Neither Samsung or Toshiba have declared any plans to create NVDIMMs directly powered by Z-NAND or XL-Flash. 451 believes this is because those two memories are not fast enough to justify the use of a DDR slot.
It remains to be seen how much demand will develop for Samsung's Z-NAND or Toshiba's XL-Flash NVMe drives. Samsung's Z-NAND drives began shipping in 2018, and 451 is unaware of any major storage system makers that are yet promoting the devices for use in their products. This situation may change over time, but we presume the current tumbling prices in the overall flash market are a headwind; they make the Z-NAND and XL-Flash devices appear more expensive than they otherwise would compared to mainstream TLC flash drives.