[Another technical post. I promise to get back to art next month.]
My last post, the one on multicore processors, ended on a down note. I don’t feel great about the future of multicore computing. However, there’s something else on the near horizon that’s going to dramatically increase the performance of photographic computer systems– the gradual replacement of disk storage with nonvolatile semiconductor memory.
Before we discuss what’s happening today, let’s get a little perspective. I’m not going to take you back to Babbage’s Difference Engine, but only about fifty years. In the mid 1950s, the standard technology for main memory was magnetic cores. Core memory provided access times on the order of the computing cycle times, was nonvolatile (you could power down the computer and it wouldn’t lose its program or main memory data, and was expensive. In 1956, IBM shipped the first disk memory system, the RAMAC 350. Designed in San Jose, and operated by vacuum tubes, the one-ton machine stored less than five megabytes (Yes, megabytes; but that was a lot; computer main memories were a few tens of kilobytes in those days). The RA stood for random access, and compared to tape, it lived up to the name, but it still had the usual disk behavior where, on a read, there’s a delay waiting for the first part of the desired data to rotate under the head, and then the data comes off fairly rapidly. The disk memory provided a storage medium midway between magnetic tape and core memory in both cost and access time. Before the disk, there was rotating magnetic memory in the form of magnetic drums, but cost per byte was much higher than for disks, and capacity was limited.
The magnetic disk was an instant success, and changed the architecture of computer systems. Large-scale transaction processing applications such as airline reservation systems were suddenly possible. IBM introduced faster, bigger, and cheaper models; competitors appeared in droves. Fast forward twenty years, to the dawn of the personal computer era. By that time, semiconductor memory had replaced magnetic cores, and that memory was volatile, that is, it forgot everything it had stored when the power went off (and whenever a cosmic ray happened by, but that’s another story). At first, magnetic disks were too expensive for personal computers, and we made do with cassette tapes (excruciating) and floppy disks (better, but still painful to someone used to minicomputers). Finally, in the early eighties, magnetic disks came to the PC, with IBM’s PC/XT, introduced in 1983 with a 10 MB disk, one of the earliest successes.
Since the RAM had become volatile, the disk offered two advantages to the personal computer: it provided a way to store data when the machine was off and automatically return it to main memory when power was restored, and it offered cheaper and more capacious, although slower, storage than the semiconductor D-RAM used for main memory. At the end of the eighties, conventional wisdom was that disks had about another ten years of life before they were supplanted by nonvolatile semiconductor memory, probably based on flash technology.
It didn’t happen that way. Giant magnetostriction, thin-film heads, perpendicular recording, and other technologies accelerated the evolution of the magnetic disk and kept it ahead of the steady advance of semiconductor memory. But it looks like time is running out for the hard disk. Already flash-based nonvolatile disk replacements are appearing in high-end laptops. In that application, their advantages are speed, ruggedness, reliability, power consumption, weight and size. In the Lenovo X300, the semiconductor “disk” can deliver up to five times better performance than the real thing. As prices come down, we’ll see much greater penetration into the portable market, until only low-end machines have hard disks.
The next step is the desktop, which doesn’t value ruggedness, weight, size, and power consumption as much as the portable market, but surely cares about speed and reliability. The advantages of semiconductor “disks” over desktop magnetic storage aren’t as great as their benefits over laptop disks, since desktop platters spin faster and their arms move more quickly. We’ll first see flash buffers on magnetic disks, then small flash memories for program swapping and access to often-used data, and then desktop computers with no rotating storage at all. Eventually, solid state storage devices will drop the interface that makes them look like magnetic disks in favor of one that is more suited to their internal structure.
We’re not talking about small performance improvements here. Using a Lenovo X300 is a startling experience. Programs load faster than you think they will. Going into and coming out of sleep state is dramatically rapid. Lightroom imports from “disk” are downright snappy. If the same level of improvement occurs when nonvolatile semiconductor memory hits desktop computing, it may be enough for us to forgive the fact that CPU utilization on our 16-processor system hardly ever gets over 30%.