idspopd - Fotolia
What’s wrong with flash storage: And what will come after?
Flash storage has taken the enterprise by storm, but its days are numbered due to an unfavourable combination of technological obstacles and manufacturing economics
Flash storage is everywhere these days. In your phone, tablet, camera, probably in your laptop and maybe your desktop too. It's certainly in your datacentre, where you'll find it in flash arrays, servers or as a cache in a rotating disk drive.
But is flash good enough for that job? Some say not, so what, if anything, could replace it?
NAND flash works by storing information in array of memory cells made from floating-gate transistors (FGT).
An FGT possesses an extra gate that is insulated and traps electrons. Multi-level cell (MLC) devices – the technology used in most flash chips – can store more than one bit per cell by choosing between multiple levels of electrical charge to apply to the floating gates of its cells. The FGT is then read by assessing the voltage level to determine whether the transistor stores 0 or 1.
So, where’s the problem? Well, flash technology has been optimised for the asymmetric usage patterns of mobile devices, where data tends to be written a small number of times and read many more.
It is also being aggressively pushed into ever-smaller form factors. Hynix announced in April that it planned to introduce 10nm NAND flash by the end of 2013. However, the smaller the form factor, the more difficult it is to make flash work well for enterprise storage.
IBM's virtual storage performance architect and "master inventor" Barry Whyte has described: "The inherent problems of switching voltages and the background work necessary to ensure clean operation at such small scales.”
He added: “You have to do more work in software to ensure durability and a 300GB [enterprise] flash drive actually has to have around 600GB of capacity to compensate for this."
According to Whyte, today's flash is "the opposite of what you want in an enterprise-class drive".
Datacentre usage patterns are quite different from that in your smartphone. Disk drives in an enterprise array are as likely to be written to as read, especially in transactional applications, which is where flash technology offers huge benefits in the form of boosted IOPS.
And a report, published in 2012 by the University of California and Microsoft Research (UCMR), The bleak future of NAND flash memory, paints a pessimistic picture of the future of flash: "Building larger-capacity ﬂash-based SSDs that are reliable enough to be useful in enterprise settings and high-performance enough to justify their cost will become challenging. We show that future gains in density will come at signiﬁcant drops in performance and reliability. As a result, SSD manufacturers and users will face a tough choice in trading off between cost, performance, capacity and reliability."
There is another generation of flash to come; triple level cell (TLC) flash, which can store three bits per cell and uses a similar technique to MLC with its multiple voltage switches per cell and that too may soon be joined by a four-bit version.
Micron and Samsung Semiconductor already sell TLC-based products. According to Micron's publicity materials, the main advantage of TLC is its low price per bit. The challenge will be, however, that higher densities combined with a greater number of bits per cell will result in poorer reliability and shorter lifetimes, according to the UCMR report.
So what can replace flash? There are some alternatives in the wings, though none has yet challenged flash's tenacious foothold, largely because the huge volumes involved ensure flash is very much cheaper than any of them. For now.
Memristors (memory resistors) (see image above) are under development, most prominently in HP's labs, and could have capacity double that of flash by the time it is productised. Memristors work by remembering a previously-applied charge in the form of a resistance when a charge in the circuit is reversed.
Williams argues that memristors are simple structures made from the same materials already in use in semiconductor foundries, and that this will lower the cost of switching from flash. The company announced in April 2013 that it planned to have the technology in commercial applications, in partnership with memory maker Hynix, by the end of the same year, but that’s been pushed back to 2014 now.
Its main claimed advantages over flash are higher density and lower power consumption.
Magneto-resistive random-access memory (MRAM)
Magneto-resistive random-access memory (MRAM) has been under development since the 1990s and uses magnetism to store data.
It works using a pair of ferro-magnetic plates separated by a thin, insulating layer. One of the two plates is a permanent magnet set to a particular polarity, while the other is a tunnel magneto-resistance (TMR) plate, whose field can be changed. Data is written and saved by re-orienting the TMR plate's field using a spin-polarised current. Because the field's resistance changes according to its orientation, an MRAM cell can be read by measuring that resistance.
Hynix, Samsung and Toshiba are developing this technology, and claim lower power consumption than flash. MRAM needs only slightly more power to write than to read, resulting in faster operation and a much-longer lifetime. They also claim added reliability because using MRAM means system designers can eliminate external components such as capacitors, resistors and batteries – keeping it simple. MRAM is already in use in Dell EqualLogic storage controllers as journal memory.
Phase-change memory (PRAM)
Phase-change memory (PRAM) is another contender, and is now being produced by Micron and Samsung. It uses heat to exploit the unique behaviour of chalcogenide glass to switch between an amorphous and a crystalline state, thereby storing binary data. PRAM is claimed to be faster than flash because single bits can be changed without first requiring that a whole block be erased, as flash does.
PRAM also degrades but much more slowly than flash, with Micron claiming an endurance of one million cycles. It is also said to be more resistant to change: flash charge leaks over time, while PRAM could hold its charge for up to 300 years – a claim that has yet to be verified. Samsung also claims the technology uses less power than flash. In 2012, Hynix and IBM announced a joint development agreement around this technology.
Racetrack memory or domain-wall memory (DWM)
racetrack memory – or domain-wall memory (DWM) – is claimed by IBM, its primary developer, to offer higher densities than flash and performance similar to that of DRAM.
This gives the non-volatile technology a unique advantage, IBM claims, namely that it could become the universal memory technology, replacing both flash and DRAM, and holding up to 100 times more data than is possible today. Big Blue showed a 256-cell prototype in 2011. Racetrack memory uses a spin-coherent current to move magnetic fields along nanoscopic perm-alloy wires. As the fields move, they pass magnetic read/write heads which alter the domains to record patterns of bits. There is no date yet for a commercial version of the technology.
Atomic memory is much further out. IBM presented in a paper in Science, January 2012, on atomic scale memory which uses just 12 atoms per bit (pictured), with an area of 4nm by 16nm, 100 times denser than a hard disk, according to Sebastien Loth, the paper's lead author. Atomic scale memory uses a scanning tunnelling microscope to perform this feat – and one would be needed in order to read the memory. Additionally, the memory would need to be kept at -286ºC to be stable. This science – arguably not yet even a technology – seems destined for the distant future.
Learn why flash cloud storage is critical if you need great performance