The solid-state drive (SSD) has actually been around for a lot longer than it's been in the consumer mindset. But only now, with its maturation and widespread use in an assortment of markets, is SSD hitting some growing pains.
Semiconductor-based SSD drives date back to the 1970s, but you'd be forgiven for thinking they came to market around the mid-2000s. In 1999, BiTMICRO introduced the first 3.5-inch SSD. Total capacity: 18GB. In 2012, OCZ, SanDisk, Samsung, and Kingston all have 500GB or higher drives, and OCZ has a 1TB drive. Price: $2,499.
In 2008, EMC was the first enterprise storage vendor to introduce enterprise SSD, and server vendors have followed suit. Just about every hardware vendor offers SSD drives in their enterprise servers.
At the time of its announcement, EMC noted that it would take 30 15,000 RPM Fibre Channel disk drives to deliver the same performance as a single flash drive, which translated to a 98% reduction in power consumption in transactions-per-second. SSD technology has only advanced from there.
SSD has settled into three comfortable niches. For the consumer, it's being adopted in laptops, particularly ultrabooks, which are trying to be as thin as possible, generate as little heat as possible, and consume as little power as necessary. SSDs fill all three of those needs.
In desktop systems, SSDs have become popular as a fast startup drive, with apps and file relegated to much larger standard hard disks. So you would use a 90GB or 120GB SSD just to boot Windows, but apps, files, videos, pictures, etc., would sit on a 2TB D: drive.
In the enterprise, the SSD has slid nicely into a position previously held by 15,000 RPM and 10,000 RPM Fibre Channel disks that provided fast access to data. In a typical enterprise setup, 15k RPM drives were the ones that sat closest to the computing side and were the most frequently accessed drives.
These drives, however, were small, rarely above 144GB in size. They used a lot of power and generated a lot of heat. The 10k drives were the next tier down in terms of importance of data. Below them were the 7,200 RPM drives, where less important or less frequently accessed data was stored.
"Servers will use ten of those 140GB 15k RPM drives at $350 apiece, short-skirt them all to use 10 percent of the capacity to get a few hundred IOPS. By doing that they use 10 drives worth of power and $3500, whereas an SSD has ten thousand IOPS and the cost is the same," said Jim Handy, principal analyst with Objective Analysis.
About the only difference is the interface. Consumer SSD remains a 2.5" drive suitable for laptop and desktop, while enterprise SSD comes on a PCI Express card. The best SATA can do in a laptop or desktop is 6Gbits/sec, while PCI Express can pump 16GB of data per second through its bus, something the enterprise needs.
Getting Too Cheap
The price of NAND flash has been plummeting, and believe it or not, this is a bad thing. "The NAND memory companies are all beginning to cut back on capacity because they feel they are in an overcapacity issue. The memory capacity is at the point where [memory] is almost free," said Bob Merritt, principal analyst with Convergent Semiconductors.
If the makers aren't making any money, then they won't stay in business long. Already there has been some shakeout, with Elpida, the last Japanese memory maker, filing for bankruptcy earlier this year. Micron Technology announced it would buy the firm.
But for now, that means bargains. Last year, NAND flash memory was about $1.60 per gigabyte. This year it dropped below $1, and hit 80 cents per gigabyte; it is continuing to drop, according to IHS iSuppli.
However, that's on the consumer side. Consumer SSD has higher error rates and shorter lifespan than the kind of quality demanded for enterprise storage, notes Michael Yang, senior principal analyst for memory at iSuppli. That price remains much higher because the enterprise uses SLC memory.
SLC, or single layer cell, means one bit of information per cell. MLC is multiple layer cells, which means two or three bits per cell. The more bits per cell, the more often it needs to be rewritten to, and that shortens cell life. This makes SLC inherently more reliable and longer-lasting, and it's why enterprise SSDs are all SLC while consumer SSDs are MLC.
So that's good news for consumers. SSD drives in the 500GB range are now selling for $400 at retailers like Fry's Electronics, Micro Center, and New Egg. Better buy them while they last, because prices will go back up eventually. These guys have to make a profit.
The Big Shrink
Intel has steadily and regularly shrunk the transistors in its CPUs for 30 years and continues to find new ways to reduce the size at an atomic level. For memory makers, however, things haven't been so easy.
A die shrink for SSD means more memory can be crammed into one chip, and thus, increase overall capacity. That's vital for SSD because it counts on increased chip capacity and more chips to raise the overall capacity of the drive.
SSD hit the mass market at 34nm, based on the already existing DRAM technology, and made a jump to 25nm. But at the 20nm process, things have hit the wall. "It's phenomenally hard to make the stuff. It's very hard to get to a 20nm process and some companies have executed better on it than others," commented Handy.
SanDisk and its memory partner Toshiba are currently at 19nm; Intel and its partner Micron is at 20nm; Hynix is at 26nm and Samsung at 27nm. In addition to the difficulty in manufacturing the memory at that small of a footprint, the rewritable memory loses integrity as it shrinks.
Handy said 35nm flash might get 10,000 erase write cycles before a cell goes bad and is no longer usable. But when you go to 25nm, you get 5,000 erase cycles and at 20nm you might get 3,000 erase cycles.
As a result, NAND flash error rates are going up significantly. What the NAND flash and SSD makers are doing, said Handy, is simply living with the defects. Instead of making a 120GB drive with 120GB of perfect memory, they make it with more than 120GB held in reserve and manage the memory through the controller. So cells go bad, reserve ones kick in, and the user always has the same amount of memory.
"There have always been error rates, but now are way higher. So instead of basic error correction schemes for 8 bits of error, they are now using very fancy error correction schemes used in telecommunications that can correct 24 or 48 bits of errors," said Handy.
Still Proving Itself
Despite the rapid maturation of the technology, SSD is still not as proven in businesses as HDD. The enterprise is extraordinarily intolerant of any kind of failures or data error, nor can it stand to have anything go down, Handy notes.
"Now you got SSDs largely from companies that don’t have a solid reputation in the enterprise. People are charmed by the speed but concerned by the lack of reputation by the companies supplying them," he said.
Seagate and Western Digital have reliable reputations after years of honing their storage technology, but they are not making SSDs. Micron, Intel, OCZ, and SanDisk are, and they are still unproven as enterprise storage makers. Seagate recently said it wants to acquire an SSD maker; OCZ Technologies is the name that keeps coming up, but both companies are mum on the subject.
"The reason Seagate is playing a minor role in SSD and WD is playing a very small role is [that] these guys want to provide SSDs that are the same quality level as their hard drives, and they are not ready to release product until they reach that quality level. Others are willing to go with pretty good levels of quality even if it's not the same quality as an enterprise hard drive," said Handy.
And that comes back to the issue of NAND flash makers trying to make things cheap, rather than reliable. "The challenge to SSD guys is the way they reduced the price so dramatically, they are relying on what a purist would call memory products with lower endurance characteristics," said Merritt.
But he doesn't think that means lower quality drives, just a slowing in the rate of growth and improvement. "We won't continue on same slope [of cost reduction] without a way to take care of errors and reduction in endurance," he said.
The one good thing about SSD is they have a wearout mechanism. "The kind of sudden, dramatic failure in a hard drive is relatively rare in an SSD. SSD has predictable failure mode. And they have hooks to read failure data, keep stats on failed blocks, and [know] at what rate they will run out of spare blocks, so you can say this drive has six months of use, this one has two years of life," said Handy.
One Year From Now…
NAND flash makers are going to have to bust through the 20nm barrier and fix those wear levels, because new technologies are coming up quickly that could replace it.
"Someone has to get past that barrier. The demand for higher performance memory, whatever the technology is, is increasing because the production of data is increasing at such a rapid rate," said Merritt.
Handy agrees. "People will upgrade from a 64GB iPhone to a 256GB iPhone when the time comes. When they do that, they are going to be getting rid of a NAND flash-based one and replacing it with something like MRAM, but they won't be aware it happened," he said.
New memory technologies like MRAM and ReRAM are coming along that will do what NAND flash does, but better. ReRAM in particular has the promise of the speed of DRAM but the non-volatile nature of flash.
These kinds of changes happen all the time. The hard drives in our PCs have gone from RLL and MFM controllers to ATA and now SATA.
Do we care? DRAM hasn't changed much but there was a time before memory DIMMs when memory expansion was done with Intel Above Boards where you had to press memory chips into a card and pray you don't bend a pin.
More recently, Intel's Xeon servers used a highly power-inefficient memory called FBDIMM, which never went into a low-power state compared to DDR2 memory. With subsequent generations of Xeon, Intel abandoned FBDIMM and moved to DDR3. Does it matter now what is in your server?
So if the iPhone 8 you buy in 2015 uses ReRAM instead of NAND flash, will you care? You shouldn't.
"I argue that something never replaces something else. What happens is new applications show up with better costs. What they do is offer new opportunities. SSD doesn't replace DRAM in servers, it offers a new spot of performance hard drives cannot fill. New technologies will do the same," said Merritt.
See also:
