TLC flash could relegate spinning disk to the background

Solid-state drives are currently blocked in their quest for dominance by cost and wear limits, but could TLC flash overcome such challenges and deal a blow to spinning disk?

Solid-state needs to move to TLC flash -- that’s triple-level cell -- to be able to deal a blow to the dominance of spinning disk in primary data stores.

Flash memory is common as muck in consumer devices such as cameras, cell phones, USB sticks and video recorders; rare in notebooks and PCs; and beginning to be the standard bearer for fast data access in storage arrays.

But what is holding it back? Why are we still dependent on fantastically complex electro-mechanical disk drives that are becoming so slow in comparison to the multi-core CPUs that are now ubiquitous?

There are two main roadblocks to flash dominance over disk drives: cost and endurance. Obviously, flash costs more per gigabyte than hard disk drives. Multi-level cell (MLC) flash is trending down towards $1 per gigabyte, whereas 2.5-inch disk drives cost between 15 and 20 cents per gigabyte.

Data on disk drives can be written and rewritten without wearing out the disk, but flash cells have a finite life. The smaller the flash manufacturing process, the shorter the working life, and the greater the number of bits per cell, the shorter the working life also.

Flash is used where its performance advantage, lower power and small size make it preferable to disk: in smartphones, tablet computers and notebooks on the one hand, and for high data access performance in servers and storage arrays on the other.

Tablet, notebook, desktop, server and storage array flash technology is currently undergoing a transition to sub-30 nanometer (nm) NAND processes, either single-level cell (SLC) or two-bit MLC technology. Sub-30 nm NAND is around 30 percent lower in cost than 49 to 40 nm NAND and we can expect a similar cost reduction from sub-20 nm NAND over sub-30 nm NAND.

These two steps forward, however, are accompanied by two steps back. The working life of small flash dies is lower, and their speed is slower. It's noticeable that fresh-out-of-the-box performance can be high but rapidly drops off as all the flash cells get used and writing new data requires full program-erase cycles.

NAND product suppliers are relying on better and better controller technology to counter these difficulties. Anobit, for example, uses signal processing technology to extract usable data from flash cells that would otherwise be considered end of life.

Assuming that controller technology can compensate for the shorter working life of sub-20 nm NAND and also of X3 (three bits per cell) NAND compared with X2 NAND, what is likely to happen when it comes to products?

The Intel-driven Ultrabook concept -- ie, a mass-produced MacBook Air-like thin notebook computer that is all-flash or hybrid -- will hopefully drive flash usage. But the initial products are quite expensive. The Asus Zenbook, for example, is $1,000 with a 128 GB solid-state drive (SSD), which is comparable to a MacBook Air, yet it does not run Mac OS and has none of that high-class Apple glamour.

If SSD costs could be lower, Ultrabooks would cost less and be more popular. There are two ways to get to that point. One is to use sub-20 nm flash process technology and so increase the number of dies per flash wafer. The other is to use three-bit MLC, otherwise known as TLC.

SanDisk -- which partners with Toshiba in flash foundries -- has been using TLC flash in more than half of its 24 nm flash production for a couple of years and has bought an SSD controller company, Pliant. It thinks there could be a transition to 19 nm NAND for its SSDs next year.

We understand SanDisk intends to bring out PCIe card format SSDs using its Pliant technology.  It seems likely that SanDisk will bring out 19 nm SLC and X2 MLC product first too. Then, when Fusion-io, STEC, Intel, Micron and others catch up with their own sub-20 nm process NAND products, SanDisk could bring out TLC flash. TLC has three bits per cell compared with SLC’s one and MLC’s two and would bring further cost advantage.

This all depends upon the flash controller technology being able to compensate for TLC NAND's intrinsically shorter working life by overprovisioning (having spare capacity set aside for use as cells die), reducing the number of writes with better ways of dealing with I/O and being able to get useful data from cells previously classed as defunct. The rewards for independent controller suppliers, such as SandForce, that do this will be immense, as they will be for SSD and PCIe flash card suppliers with their own controller technology: Intel, Fusion-io, Micron, OCZ with Indilinx, SanDisk with Pliant, and STEC.

Things are on course for NAND to become, say, only three to five times more costly than disk per gigabyte, compared with the current ratio of about 10:1. That could trigger the emergence of large flash memory tiers of storage in servers and PCs, with vast increases in data access speed to an entire application's working set held in solid-state memory. Disk would become, literally, a backing store. Isn't it amazing what a little TLC can do?

Chris Mellor is storage editor with The Register.

Read more on Storage fabric, switches and networks