Communications speeds _ALSO_ fit with Moore's Law, interestingly enough
Michael Gat wrote:
<<It also made it clear that he sees something that you don't. Namely that the speed of the processor is no longer usually the bottleneck, and that MHz are no longer the key criteria. Fast disk drives and an upgraded network make a much bigger difference in real world performance.
Simply put, all the stuff around the processor has just not kept up with Moore's law. All that stuff needs to catch up before the processor speed makes a big difference in real world applications.>>
Let's look at where things were 15 years ago, in 1986. This is a good year to look at, because the industry was in a recession, because Intel was working hard to release the 386, because Apple was starting to do very well with the Mac, and because 15 years is just a good baseline for comparison. (Others can do the calculation for 1991-2001.)
* A typical processor speed was about 6-12 MHz, in about 30-70,000 transistors.
* A typical amount of DRAM was 256-640 KB. SRAM cache was limited to on-die cache, and not much of it.
* A reasonable-to-large disk drive for a PC was 20-40 MB in 1986..and I paid $1400 for an 80 MB "whopper" in mid-87.
Fifteen years later:
* A typical processor speed is 1 GHz, in about 10-50 million transistors. But handling chunks of data a few times larger and with pipelines and all the performance-enhancing tricks. A 1000x enhancement is a reasonable estimate. (Benchmarks over the period would give a more accurate figure, but the transition from around a MIPS to a GIPS is about right. Floating-point performance has gone up more, as extra transistor count is devoted to FP.)
* A typical amount of DRAM is about 100-500 MB, for roughly a 1000x enhancement. (This carries over to more than just PCs. My Mac Plus in 1986 had 1 MB of RAM, while the machine I am typing on right now has 576 MB of RAM.) Access time is a little bit faster, but not by more than a factor of several.
So far, it's looking like "1000x in 15 years" is about right. Let's now see what disk drives have done:
* A typical disk drive is 30-80 GB, for--drum roll!!--roughly a factor of 1000x.
(Seek times are sub-10 msec now, and data are moved in bigger chunks.)
Factoring in the fudge factors for pipelining and other processor tricks, for faster access times, the 1000x raw figure is probably closer to "5000x" for _each_ of the three.
Factoring in the price drops (e.g., a PC-AT in 1986 was $3000 or more, a disk drive was $500, memory was also expensive), I'd say "10,000x" is a rough measure for overall improvement.
All in all, it looks like disk drives _have_ kept up with processor speed.
How about communications? Obviously not, if measured by average connect speed. But there are some gotchas...
* In 1986 a typical home user had maybe a 300-baud acoustic coupler modem (Radio Shack, for example). If they had a modem at all. Some had 1200-baud modems. (My own first modem was 2400, in 1987. Then the jump to 9600, then to 33 KB, then "nominal" 56K.)
* In 2001, most users have at least 48K-baud, depending on the quality of their phone lines. Some have ISDN, some ADSL, some cablemodem, some satellite dish, some even have their own T3s (friends of mine in the Santa Cruz mountains). For the modem users, the enhancement has been about 50x. For the others, upwards of--drum roll again!!--1000x. Gee, this number keeps popping up.
And the "gotchas" I mentioned are that connecting the "last mile" is only part of the story: there's so much _more_ to connect to out there than there was in 1986.
Another gotcha is that users typically only need to connect for short periods of time. Unless one is downloading DIVXed movies (Napster-style trading in DVDs), the extra performance of a 1000K-baud connection over a 200K-baud connection is not seen.
(Being out in the country, I am beyond reach of ADSL at this moment. And no cablemodem. Yet I do quite well, thank you, on my 56K dial-up, which is usually limited to 48K. My cablemodem friends can download files perhaps 20x faster, but how often?)
In summary, I didn't set out to generate this "1000x" number that keeps popping up (a bit more if performance fudge factors are taken into account, and even more if average selling prices are also factored in).
But it sure looks like the oft-cited point that "communication speeds and disk drive capacities have not kept up with processor speeds: is simply not so. Disk drives have more than kept up with processor speeds (though disk drive speeds have not, and will not, for physical rotation rate reasons). Communication speeds have gone from roughly a kilobaud in the mid-eighties to about 50x that for dial-up users, 500x that for most ADSL users, and 1000x that for cablemodem users. Optical fibers, which are starting to be deployed, will likely push communication speeds way out in front of processor speeds.
I enjoyed doing this calculation, so, thank you, Michael Gat!
--Tim May |
