Flash could be CHEAPER than SAS DISK? Come off it, NetApp A NetApp pitch at the Flash Memory Summit said the cost-per-GB price trends for cheap TLC flash mean it will dip below SAS disks in 2017. TLC is triple-level cell flash which holds three bits per cell, half as much again as the standard 2-bit or MLC NAND that is used in SSDs and server flash cards. Because TLC is denser in …
> SAS is going to remain cheaper for a while.
Unless we manage to get our data volume requirements down.
Most disks in enterprise arrays are small and fast because speed > capacity. That's why we have SAS with SCSI rather than SATA. Can we use smaller SSD arrays because they have extra speed?
This seems a bit FUDDY. Internet retail prices of MLC flash drives have already fallen below SAS HDD OEM pricing.I don't know why anyone would use enterprise HDDs.
As to the supply issue. Pricing tells the story. If there were a supply problem, pricing would be going up not down. I agree that flash won't replace bulk disk drives with 5TB capacities any time soon, but enterprise "fast" drives are dinosaurs. There's plenty of flash production capacity for that segment to be satisfied. Note that the replacement rate of flash for enterprise HDD is significantly lower than 1:1 for both drives and total capacity. The SSD are much faster, and don't need short-stroking.
You can't buy 5 TB flash drives, at least not at the same price per GB as you can buy sub 1 TB flash drives. So even if they cost less per GB than a 5 TB SAS drive, you'll need more/bigger Netapps.
Looked at that way, it kinda makes sense that Netapp would try to get people to focus on the cost of the drives, not the array they're going into :)
-Disclosure NetApp Employee, though not an official NetApp response, more of a personal rant -
@Anonymous - the $/GB of SAS increases relative to the $/GB of SATA, i.e. it's not going to fall at the same rate, e.g. SATA might drop around 20% vs SAS at about 5-10%. If the system was in $/GB you'd see a bunch of logarithmic curves with TLC flash dropping really fast and crossing over the curve for SAS but still being an order of magnitude more expensive than SATA. It might seem counterintuitive initially but this is actually one of the clearest ways of showing the relative costs.
@elReg - the "Enterprise" segment for HDD's that goes into arrays is about 1/10th of the overall HDD market, most of which goes into mobile devices today, though that too may end up being primarily flash at the rate things are going. It would be entirely possible for the existing NAND capacity to completely replace the enterprise disk segment even without new investment in fabs, which is likely to happen in any case even if it's just to satisfy the insatiable demand for low power mass storage in mobile devices.
@Jim - Enterprise devices have significantly lower failure rates in the field than run-of-the-mill OEM SAS/SATA, if you're on the hook for replacement under warrantee, then it makes economic sense, not to mention avoiding RAID reconstructs is a really good idea if you care about consistent performance in a shared environment.
@MyffyW & TwoEyedJack - Disk MTBF is a really bad predictor of Raid-set and array reliability (theres an IEEE paper I can point you to with lots of math and talk of Weibull and Poisson failure distributions and dependent vs in dependent failure data). Also the annual failure rate for enterprise class drives is significantly better than the stuff google was reporting on. Having said that, from what I've seen, eMLC drives are showing incredibly good reliability figures in the field.
SATA Disks will probably be replaced by something else for mass storage eventually, but probably not before any of us really care any more, but my personal bet is that it won't be NAND "as we know it" which is reaching the end of it's density journey over the next 5 years if the research is right, though you can make it denser, the speed will go down to almost disk like levels of IOPS / GB for writes. Having said that I know a bunch of smart people who disagree with me.
Regards
John
"@Anonymous - the $/GB of SAS increases relative to the $/GB of SATA, i.e. it's not going to fall at the same rate, e.g. SATA might drop around 20% vs SAS at about 5-10%. If the system was in $/GB you'd see a bunch of logarithmic curves with TLC flash dropping really fast and crossing over the curve for SAS but still being an order of magnitude more expensive than SATA. It might seem counterintuitive initially but this is actually one of the clearest ways of showing the relative costs."
Thanks for responding to my question, I'd really like to see the non-normalized version of this chart as otherwise it reminds me of a book I own called 'How to Lie with Statistics' as where the choice of axis and normalization is made to reinforce the premise rather than make it clearer.
Wow, I'm sure all of the flash folks would love to be able build a new fab for a billion dollars. But the cost for a large enough fab to make economic sense is more like $5-6b. At a minimum. Samsung estimated that the investment in their new fab in Xi'an will be $7b. That will probably prove to be low. 3D production has more steps, is more complicated and costs more than planar NAND to build.
What that last paragraph ignores, IMO, is that today there will only be 4 companies (at the most--Samsung, Toshiba/Sandisk, Micron, Hynix) building a new fab, not a dozen or more as in the Bad Old Past and those companies have learned how to bring the capacity up in stages rather than all at once. And the beauty of the SSD market for the flash companies is that they can get rid of extra flash by making larger capacity drives for not very much more money and still make good profits. At least, they can if they don't really drastically misjudge the market. There is no incentive for a company to overbuild and over-supply the market with only a few relatively strong competitors as there was when the market was more fragmented and had many weak competitors that could be picked off by lowering prices.
I know what this sounds like, but -- "This Time Really Is Different". Really. At least for the NAND vendors that have successful SSD programs. At least until SSDs saturate the market. And, as Rakers and Mellor point out, we are a long way off from that--it will take many years and many billions of dollars of investment before that can happen (sigh of relief here from Seagate and WD). As the only pure play in NAND, Sandisk in particular ought to do well in the coming years, especially given the way they have been building up their SSD business (disclosure: yeah, I own the stock; but no, I am not employed by them or by any company with any interest in them). If Micron and Hynix can get their NAND acts together, they should also do well, although Hynix in particular has legal problems ahead of it and currently badly lag.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
