Tuesday, March 21, 2006

Bits and More Bits

I have spent a considerable part of my career working in the general area of digital storage, on software related to hard disks, digital tape systems, CD-ROMs, etc.. The computer industry calls this entire area of technology simply “storage", by which (generally) they mean non-volatile secondary storage. More informally, storage is all those places your computer more-or-less permanently saves its bits, whether or not the computer’s power is on.

Virtually all aspects of storage have changed dramatically in the 30 odd years of my career, but clearly the most significant changes have been in hard disks. In the late 1970s, I purchased the first hard disk I ever owned. It was a used Memorex 630, manufactured in the late 1960s; I paid $10,000 for it. This marvel was the size of a extra-large washing machine, and had just over 7MB of storage. That works out to 0.14 cents per bit — very cheap at the time. I spent hundreds of hours designing and building a controller to interface that disk to my Z80 CP/M system, for the sole purpose of speeding up my software development process. By today’s standards 7MB is laughably small, but in the late 1970s that system gave me (so far as I know) the biggest, baddest microcomputer-based development system in the world. At that time, “serious” software development for microcomputers was mainly being done on minicomputers, and part of the reason for that was the better storage found on the minis.

Just recently I had to run down to Fry’s Electronics to buy a hard disk as a repair part for one of my servers. The disk that broke down and needed replacement was 40GB — over 5,000 times the capacity of that old Memorex 630. And it wasn’t the size of a washing machine — it was more like the size of a paperback book. But Fry’s didn’t have any disk drives as small as 40GB! The smallest one I could buy was 120GB — but hey, it was only $89, so what the heck. The cost per bit on that disk drive was about 0.000000074 cents per bit — almost two million times cheaper than that used Memorex 630.

Hard disks keep getting cheaper and cheaper, especially in terms of price per bit of storage (because they keep getting bigger as well). Over the thirty years since I bought that Memorex 630, the price per bit for storage has dropped at a quite consistent rate, with storage each year costing about 60% of what it cost the year before. If you project forward ten more years, a 10TB (10,000GB) hard disk should cost about $42 — except you probably won’t be able to buy one that small!

But will hard disks still dominate in the future?

Hard disks currently have the advantage in price per bit, and they are reasonably fast. But the rate of advance in the speed of hard disks has not kept pace with the increase in speed for every other aspect of computer systems. The result of this divergence is that hard disks are today the major bottleneck in computer system performance. The fact that your PC takes a minute or so to boot is almost entirely due to the low speed (by comparison to the rest of your computer) of the hard disk. A modern computer can retrieve an arbitrary chunk of data (a word, it’s called) from its main memory (RAM) in about one billionth of a second — but to do the same thing from a hard disk takes several thousandths of a second, which sounds fast until you realize that’s millions of times longer than from main memory.

So hard disks are the bottleneck. What kind of storage is faster than a hard disk? Several, actually, but only one is currently a serious contender: FLASH memory (which comes in several technical variations). The price per bit today for FLASH memory is about 100 times higher than hard disks — but it is falling much faster. The price per bit may cross that of hard disks as early as five years from now, and almost certainly by ten years. And FLASH memory is superior in just about every performance category: it is much faster than hard disks, much smaller, uses less power, and has no moving parts. There’s one big difference with FLASH, though, that points to an area where hard disks may remain cheaper: the cost of making FLASH is essentially linearly related to its storage capacity. This is not true for hard disks — it costs very little more to add more bits to a hard disk, once you have paid for the basic mechanism.

Tom’s prediction (barring the introduction of some radical new technology): ten years from now, FLASH memory will dominate storage, with hard disks relegated to the particularly large, and probably tertiary, storage systems. Old-timer storage geeks like myself will recognize this as being parallel to the old paradigm of RAM-disk-tape, wherein RAM was used for primary storage, disk for secondary, and tape for tertiary. I think we’re going to see tape completely relegated to backup (actually that process is already well underway), with disk becoming the new tertiary media, and FLASH the new secondary. If I’m correct, that means production volumes of FLASH will go up (driving price per bit down even more). The opposite will be true with hard disks: volumes will go down, at least reducing the rate by which the price per bit drops.

Given the rate of innovation in the storage industry, it’s always possible that some new storage technology will pop up and displace both FLASH and disk. Possible, but not, I think, very likely. The history of storage has been mostly one of incremental advances, with fundamental new technologies coming along relatively rarely. In fact, nearly all the contenders for mass storage have actually been based on the same technology (magnetic storage): hard disks, tape, core memory, and even bubble memory, are all variants of magnetic storage technology. FLASH is the first semiconductor contender, with battery-backed DRAM another variant that might gain some traction. Those two fundamental technologies underlie every commercially significant storage technology for the last forty years. Of all the experimental technologies I’ve read about or heard of, the only one that seems particularly promising to me is three-dimensional optical storage (especially holographic, but also the non-holographic). The reason I find them promising is the extraordinary densities they could theoretically achieve. But these are very much laboratory curiousities at the moment, and I wouldn’t give them too much of a chance of emerging into high-volume production within ten years.

Bet on FLASH.