The status of NAND flash memory now and in the near future
Flash memory chips have driven a revolution in storage, and faced a recent shortage because of huge demand. That's easing, and advances in NAND technology show a great future.
Under the hood, NAND flash memory has been in a major transition for the last year. Flash is a fiendishly complicated balance of physical properties, compounded with production processes, data integrity requirements and write durability questions. In a nutshell, the original flash approach (known as 2D NAND) was running out of headroom as we headed into 2016, with no way to keep increasing capacity per die by shrinking features.
Enter the concept of 3D NAND, where flash cells are stacked in the third dimension. Capacity per die increases of 48x looked achievable with relatively little effort, and this increase would reduce cost per bit perhaps as much as 4x once in high-volume production. The difference in capacity increase and price reduction ratios stems from the fact that many of the production processes sputtering each cell layer on the die still occur with 3D NAND.
So much for the plan
It turned out things didn't go according to the plans of the NAND flash memory chipmakers. One of the first casualties was the ease issue. Maintaining vertical alignment of all those layers proved very challenging.
Permissible tolerances of a vertical stack are approximately 1 degree for a 48-cell stack, and it gets worse with more layers. This caused something of a rethink of the whole concept of 3D NAND.
Chipmakers have migrated to an approach that creates multiple 48- or 64-layer stack blocks on a die, placing each of these on top of the others, but using a coarsely sized vertical interconnect access to connect the stack blocks. This allows for a much greater tolerance for the vertical alignment of layers.
The changeover to 3D NAND has taken longer than planned, resulting in a shortfall of die as factories have gone through conversion to the new processes.
The net effect is that we have had announcements of both 64 GB die and 1 TB die in the last few months, reflecting these two approaches. The 1 TB die is likely to see even further stacking, and single chips with 4 TB, or even 8 TB, are likely within 12 months or so. With these devices, 100 TB 2.5-inch SSDs announced at the September 2016 Flash Memory Summit are looking very real in 2018.
The alignment issues and change in technical approach have affected the market significantly. The changeover to 3D NAND has taken longer than planned, resulting in a shortfall of die as factories have gone through conversion to the new processes.
More than misalignment
Alignment isn't the only issue to hinder 3D advancement. Clearly, triple-level cell (TLC) flash is the attractive spot in NAND flash memory capacity, giving 3x the original capacity per cell.
NAND devices are quasi-static in their properties over time, with the result that the separation in voltage levels detectable in a cell shrinks, bringing cell state voltages closer to each other. This increases misread rates, and is more pronounced in TLC, with eight distinct states, compared with the two states in the original single-level cell flash.
This shrinking requires countermeasures, including resetting thresholds over time, extending error detection/correction codes and adding very sophisticated retry procedures with threshold and timing adjustments. This is heavy science (and math), and getting it right, followed by an economic implementation, has had its challenges. The positive results of the effort have been major increases in die yield and write durability.
Even so, we will see a family approach to SSDs, where members of a family with essentially the same construction will be differentiated by read/write speed, write durability and temperature range. The ranges will be significant, too, with as much as a 10x spread in daily drive writes from least to most, for example.
The result of the process delay has been a reduction in flash die availability, as NAND flash memory fabs were idled to enable the changeover. This hit hard in the second half of 2016, with as much as a 10% increase in SSD prices -- and it's continuing. The problem was compounded by rapidly increasing demand for SSDs, due to the realization that the enterprise HDD is becoming an obsolete concept.
Further stress was added by Samsung's failure, recall and attempted replacement of the Galaxy Note, which pulled more than 100 million die out of the market.
Relief is approaching
The second half of 2017 will bring relief. The technology issues of stacked 64-layer die seem to be behind us, and by mid-year, single stack die will be shipping in volume. Additional fab capacity will be coming online from dynamic RAM and NOR flash fabs repurposed to 3D NAND, and from new fab capacity. By year-end 2017, prices should be falling back to pre-crisis levels and below; in 2018, we should see SSDs closing toward price parity with HDDs.
It's worth commenting on what this has meant for the HDD market. Faced with declining demand in 2015 and 2016, HDD production capacity has fallen considerably, so HDDs aren't well positioned to take back market share. In addition, the large capacity HDDs that are in demand aren't well suited to primary storage use, being shingle write technology products that are relatively slow -- especially when writing -- and totally unsuitable for operating system drives, nixing them for PCs and many server uses.
The industry looks to be heading for an all-flash future, beginning in 2018 with 30 TB to 100 TB bulk drives being the sweet spot for secondary storage.
Expect some debate about using quad-level cell (QLC) NAND flash memory for storage instead of the soon to be ubiquitous TLC. Based on all of the research investment into physics and data dynamics for TLC, applying the same thought processes to QLC, with 16 possible cell states, just seems like a logical follow-on, pointing to even further cost reductions and capacity increases.
The advent of ubiquitous, very fast storage will profoundly change IT. Servers will do much more, reducing server sales, for example, while drives with 10 million or more IOPS (announced at the Flash Memory Summit) will revolutionize hyper-converged offerings.
