Tuesday, October 20, 2015

Storage Technology Roadmaps

At the recent Library of Congress Storage Architecture workshop, Robert Fontana of IBM gave an excellent overview of the roadmaps for tape, disk, optical and NAND flash (PDF) storage technologies in terms of bit density and thus media capacity. His slides are well worth studying, but here are his highlights for each technology:
  • Tape has a very credible roadmap out to LTO10 with 48TB/cartridge somewhere around 2022.
  • Optical's roadmap shows increases from the current 100GB/disk to 200, 300, 500 and 1000GB/disk, but there are no dates on them. At least two of those increases will encounter severe difficulties making the physics work.
  • The hard disk roadmap shows the slow increase in density that has prevailed for the last 4 years continuing until 2017, when it accelerates to 30%/yr. The idea is that in 2017 Heat Assisted Magnetic Recording (HAMR) will be combined with shingling, and then in 2021 Bit Patterned Media (BPM) will take over, and shortly after be combined with HAMR.
  • The roadmap for NAND flash is for density to increase in the near term by 2-3X and over the next 6-8 years by 6-8X. This will require significant improvements in processing technology but "processing is a core expertise of the semiconductor industry so success will follow".
Below the fold, my comments.
  • As I've written before, tape recording technology tends to lag hard disk by about 8 years, so the tape roadmap out to 2022 is very credible. But even though 48TB/cartridge is impressive, it may not matter. Tape is losing market share in cold storage for reasons other than raw cartridge capacity, and with media vendors dropping out, and system vendors down to three, customers have to be increasingly concerned about its long-term viability.
  • The optical roadmap seems less credible. There are no dates, there are significant physical problems, and because it represents a significant increase over the historical 12%/yr density increase unless the roadmap extends through the next decade.
  • My friend Dr. Pangloss would enjoy the hard disk roadmap. HAMR was supposed to ship in 2009. Six years later, it has yet to ship in volume. Everyone understands that the HAMR-BPM transition will be even harder than PMR-HAMR. The idea that, after being six years late, HAMR will be in the market for only six years before being supplanted by BPM only six years from the proof-of-concept stage is positively Panglossian.
  • The cross-section in the slides of 3D NAND flash is mind-boggling. On the other hand Fontana is right that betting against the semiconductor industry being able to get process technology working has a poor track record. So this roadmap is fairly credible.
So far, all this is in terms of bits per square inch. But what we're really interested in is $/GB. Daniel Rosenthal's research suggests that, historically, the increase in bits per square inch for hard disk accounted for about 3/4 of the decrease in $/GB. There are reasons to believe that the factors accounting for the remaining 1/4 are now much less effective, so $/GB should track density more closely.

One fascinating slide in Fontana's presentation includes for tape, flash and hard disk the total revenue for a year divided by the number of gigabytes shipped.  Extracting these numbers from the slide, and dividing the $/GB for flash by the $/GB for hard disk gives us the cost ratio and the following table:
  • Year  Cost Ratio
  • 2008 12.2
  • 2009 13.1
  • 2010 17.7
  • 2011 11.6
  • 2012 7.8
  • 2013 8.7
  • 2014 8.4
As you can see, despite the disk industry's troubles with floods and new technology, the $/GB of flash has remained about an order of magnitude greater than that of hard disk. It is a more valuable product, so even if there were no supply constraints, it would command a higher price. But there are supply constraints, as I pointed out in Another good prediction. So hard disk will continue to be the medium on which bulk data lives and dies.

The roadmaps for the post-flash solid state technologies such as 3D Xpoint are necessarily speculative, since they are still some way from shipping in volume. But by analogy with flash we can see the broad outlines. They are a better technology than flash, 1000 times faster than NAND, 1000 times the endurance, and 100 times denser. So even if the manufacturing cost were the same, they would command a price premium. The manufacturing cost will initially be much higher because of low volumes, and will take time to ramp down.

No comments: