## SSD and HDD capacity goes on embiggening

### But is it better?

**Comment** SanDisk and Sony have announced a roadmap for Memory Stick products that take it from its current (Memory Stick PRO) 16GB capacity up to 2TB. Do we have a tape-like generational roadmap here?

LTO is a classic example of tape generations with LTO 4 at 800GB native capacity being the current generation and LTOs 5 and 6 ahead, if all goes well, with 1.6TB and 3.2TB. The general rule was for a doubling of tape capacity with each generation and, hopefully, a doubling of I/O speed as well, with two years or so being the generation gap.

The capacity and speed increases came about through better chemistry for the tape's recording layer and better electronic technology for the read/write heads.

We have a generational increase in disk capacity as well with successive generations of perpendicular recording, and whatever follows it, steadily increasing HDD areal density and thereby capacity. Now it seems clear that a similar process will occur with NAND flash. It will be driven by different factors from those driving tape and disk capacity increases, though.

### NAND development axes

There are two development axes as it were - the first is the NAND cell bit density, and the second the process feature size.

Bit density is a measure of how many binary stores there are in a flash cell. A single level cell (SLC) has 1 binary digit. A multi-level cell (MLC) has more with, for example, SanDisk announcing 2-bit cell (2X) products now and 3- (3X) and 4-bit (4X) product coming. If a 1-bit cell chip holds 16Gbits then a 2-bit one will hold 32Gbits, a 3-bit one 48Gbits and a 4-bit one 64Gbits.

A by-product of increasing cell bit density has been a slowing of the I/O write rate with SLC NAND being the fastest and each increase in bit density slowing NAND down. For example. Intel's SLC X25-E SSD does sequential reads at 250MB/sec and sequential writes at 170MB/sec. The MLC X25-M sequentially reads at 250MB/sec but sequentially writes at 70MB/sec, less than half the speed of its SLC sibling.

A second by-product of increased bit density is reduced write endurance. This is being countered by adding parallel channels to controllers and avoiding unnecessary erase-delete cycles when writing data by over-provisioning the NAND and by clever file systems such as SanDisk's ExtremeFFS. This is a uniquely flash phenomenon; it doesn't apply to tape or to hard drives.

Another by-product of increasing cell bit density is to lower the cost per bit of the flash. It's a case of what you lose on the speed and endurance swings you gain on the capacity roundabout.

Process feature size refers to the size of individual NAND cells. The smaller they are the more of them you can fit onto a 300mm NAND flash wafer. That cuts the cost per bit. The smaller they are the more cells you can have in a flash product, such as a Memory Stick, USB thumb drive or solid state drive (SSD). It's reckoned that a move from 56nm feature size to 43nm effectively doubles NAND capacity at any product format. Thus a 32GB SSD with a 56nm feature size will become a 64GB SSD with a 43nm feature size.

There isn't a standard process size at each generation. Toshiba and SanDisk, who jointly operate flash fabs, are moving to a 43nm process size from a 56nm one. Intel and Micron, participants in Intel Micron Flash Technologies (IMFT), are moving from a 50nm process to a 34nm process, bypassing any process in the 40-49nm area. It's expected that SanDisk and Toshiba will progress to a sub-40nm process next.

What looks to be happening is that a given process size starts at 1 or 2-bit and then progresses to 3-bit and 4-bit cell densities. It's expected that the SSD suppliers will be able to maintain an average capacity increase rate by tweaking process size, bit density and chip packaging parameters over time.

### Capacity increase rates

The SanDisk/Sony Memory Stick PRO roadmap goes from 16GB maximum capacity to 32GB and then on to 2TB. If we assume a capacity increase rate of 50 per cent a year on average that gives us the following timetable:

2009 - 32GB

2011 - 64GB

2013 - 128GB

2015 - 256GB

2017 - 512GB

2019 - 1TB

2021 - 2TB

Let's apply the same increase rate to SSDs starting from a 2.5-inch form factor 512GB product (Toshiba) today:

2009 - 512GB

2011 - 1TB

2013 - 2TB

2015 - 4TB

2017 - 8TB

2019 - 16TB

2021 - 32TB

How does this compare to expected HDD capacity increases?

Toshiba has announced a 500GB 2.5-inch HDD late last year, joining other suppliers such as Seagate. Lets have a start point of 500GB in 2009 and, applying a 50 per cent per annum capacity increase rate, see what we get:

2011 - 1TB

2013 - 2TB

2015 - 4TB

2017 - 8TB

2019 - 16TB

2021 - 32TB

We get parity, ignoring that slight difficulty of a TB being 1024GB in one table and 1000GB in the other. So let's say we get rough parity. That will make life interesting.

Of course this destination point is erected on an assumption of 50 per cent per year capacity increases with both SSD and HDD technology. That could be wrong. HDD suppliers, feeling the heat of SSD pursuit, could aim to raise capacity faster than this. The flash suppliers may run into technology difficulties. Who knows? We're guesstimating here.

Also there isn't one SSD capacity increase curve but several, one for each cell bit density level.

The picture is complicated by cost/GB differences between the two technologies and price decline rates as well as by I/O speed and write endurance differences. (If anyone can offer intelligence about the comparative price decline rates for HDDs and SSDs, that would be very interesting to know.)

One conclusion is that it is going to become harder to distinguish the right random-access storage technology to use in particular applications, choosing between, say, differently priced and capacity level 2.5-inch HDDs with SATA or SAS interfaces and 5,400rpm, 7,200rpm, and 10,000rpm spin speeds on the one hand, and differently priced and capacity level 1X, 2X, 3X, 4X and maybe even 5X 2.5-inch SSDs on the other, with various speeds and write endurance levels - not to mention hybrid devices with thumping great NAND caches in front of their spinning platters.

Storage life is not going to get any simpler. ®