Latest forecast question is: Will 3D TLC NAND be GA’d in major storage products in 12 months?
Splitting up the QoW into more answerable questions:
A) Will any vendor be shipping 3D TLC NAND SSDs/PCIe cards over the next 9 months?
Samsung will is reportedly already shipping 3D TLC NAND SSDs and PCIe cards as of August 13, 2015 and will be producing 48 layer 256Gb 3D TLC NAND memory soon. Unclear what 3D TLC NAND technology will be shipping in the next generation drives due out soon but they are all spoken of as read-intensive/write-light storage.
One consideration is that major storage vendors typically will not introduce new storage technologies unless it’s available from multiple suppliers. This is not always the case and certainly not for internally developed storage but has been a critical criteria for most major vendors. But in the above reference, it was reported that SK Hynix and Toshiba are gearing up for 2016 shipments of 48 layer 3D TLC NAND as well, how long it takes to get these into SSD/PCIe cards is another question.
A number of startups are rumored to be using 3D TLC and Kamanario has publicly announced that their systems already use 3D TLC.
My probability of a second source for 3D TLC storage coming out within the first 9 months of next year is 0.75
B) What changes will be required for storage vendors to utilize 3D TLC NAND storage?
The important changes will be SSD endurance and IO performance.
NAND endurance is rated at DWPD (drive writes per day). Current Samsung 3D TLC SSDs are reportedly rated anywhere from 1.3 to 3.5 DWPD for a 5 year warranty period and newer 3D TLC SSDs are rated at 5 DWPD (unknown warranty period). Current enterprise (800GB) MLC drives are reportedly rated at 10-25 DWPD (for 5 years). So if we use 3.5 DWPD for 3D TLC and 17.5 DWPD for MLC, 3D TLC NAND has a ~5X reduction in endurance.
As for performance, if we use the Samsung reported performance of 160K random reads and 18K random writes vs. an HGST 800GB MLC SSD that has 145K random read and 100K random write performance. There is a reduction of ~5.6X in write performance. Read performance is actually better with 3D TLC NAND.
In order for major vendors to handle, a reduction in 3D TLC endurance, they will need to limit the amount of data written to these devices. Conveniently, in order for major vendors to deal with the reduction in 3D TLC write performance, they will also have to limit the amount of data written to these devices.
Hence, one potential solution is a multi-tiering, all flash array which uses standard MLC SSD/PCIe cards to absorb the heavy write activity and data from this tier, that is relatively unused, could be archived (?) over time to a 2nd tier of storage consisting of 3D TLC SSD/PCIe cards.
This is not that unusual and it’s being done today for hybrid (disk-SSD) systems with automated storage tiering. Only in this case, data is moved to SSD only if it’s accessed frequently. For 3D TLC the tiering policy should be changed from access frequency to time since last access. Doing so in a hybrid array with disk, MLC SSD and TLC SSD, would require the creation of an additional pool of storage and could be accomplished with software changes alone. There are current major vendor storage systems that already support 3 tiers of storage. And some which already support archiving to cloud storage, so these sorts of changes are present in current shipping product.
So yes there’s a reduction in endurance and yes it has worse write performance but it’s still much faster than disk and most major vendors already have software to be able to handle diverse performance storage. So accomodating the new 3D TLC storage shouldn’t be much of a problem.
New storage technology like this usually doesn’t require a hardware change to use. So the only thing that needs to be changed to accomodate the new 3D TLC is software functionality
So if the 3D TLC 2nd source was available there’s a 0.9 probability that some major storage vendor would adopt the technology over the next year.
3) What are the advantages of 3D TLC storage?
Price should be cheaper than MLC storage and the density (GB/volume) should be better. So in this case, it’s a reduction in cost/GB and increase GB/volume. So for these reasons alone it should probably be adopted.
The advantages are good and would certainly give a major vendor an edge in capacity density and in $/GB or at least get them to parity (barring any functionality differential) with startups adopting the technology.
So given the advantages present in the technology, I would say there should be a 0.7 probability of adoption within the next 12 months.
Forecast for QoW 15-002 is:
0.75*0.90*0.70 = 0.47 probability of YES adoption or .53 probability of NO adoption of 3D TLC NAND in major storage vendor products over the next 12 months
I have an update to my post that forecast for QoW 15-001 as a No with 0.62 probability. This question was on the adoption of 3D XPoint (3DX) technology in any enterprise storage vendor product within a year.
It has been brought to my attention that Intel mentioned the cost of producing 3DX was somewhere between 1/2 and 1/4 the cost of DRAM. Also, recent information has come to light that Intel-Micron will price 3DX between 3D NAND and DRAM. So my analysis as to the cost differential for caching technologies is way off (20X). So there would be a significant cost advantage in using the technology for volatile and non-volatile cache. Even if the chips cost nothing, it might be on the order of $3-5K cheaper with 3DX than battery/superCap backed up DRAM and volatile DRAM caching. So it exists but less than a significant cost saver. So this being the case, I would have to adjust my 0.35 probability of adoption in this use up to 0.65. I failed to incorporate this parameter in my final forecast, so all that analysis was for nothing.
Another potential use is as a non-volatile write buffer for SSDs and even more important for 3D TLC NAND (see above). As this is in an SSD, software and hardware integration is commonplace so there’s a higher probability of adoption there as well. And as there are more SSDs than DRAM caching the cost differential could be more significant. Then again, it would depend on two technologies being adopted (TLC and 3DX) so it’s less likely than any one alone.
The other news (to me) was that Intel announced they would incorporate proprietary changes in DIMM bus to support 3DX as one approach. This does not lend credence to widespread adoption. But probably only applies to server support for the technology, so I would reduce my probability there to 0.55
Updated forecast for QoW 15-001 is now:
- chip in production stays at .85, so there’s still 2.6 potential systems that could adopt the technology directly
- 0.85 probability that chips in production * 0.55 probability of servers with the technology * 0.65 probability that a storage vendor would adopt the technology to replace caching, so (=) ~0.30 probability of server adoption in storage, and with 18 potential vendors thats another 5.5 systems potentially adopting the technology.
- Add in the two-three startups that likely will emerge, with similar probability of adoption, or 0.30, which is another 0.9 systems
For a total of 2.6+5.5+0.9=9 systems out of ~24 or 0.38 probability of adoption.
So my updated forecast still stands at No with a .62 probability.
Question of the month (QoM for February is: Will Intel Omni-Path (Architecture, OPA) GA in scale out enterprise storage by February 2016?
First, let me state that QoM stands for Question of the Month. Doing these forecast can be a lot of work, and rather than focusing my whole blog on weekly forecast questions and answers, I would like to do something else as well. So, from now on we are doing only one new forecast a month.
