IanG on Tap

Around a year ago, I wrote a blog entry bitching about (amongst other things) the fact that if you wanted a fast laptop hard disk, the fastest choice was limited to 60GB, and that this had been the case for over a year.

Finally, Seagate seem to have introduced the hard disks they preannounced about 10 months ago. Whether they're actually available remains to be seen. As you can see I was hoodwinked by their June 2004 announcement into thinking that the hard disks might be available some time soon. So this time round I will remain sceptical until I have one inside my laptop.

Unseemly Gloating

I like to say "I told you so". :-)

Looking back at my 'end of Moore's law' article in which I discussed the lack of progress on laptop hard disks, I acknowledged the possibility that I might be mad to predict that Moore's law grinding to a halt. This has been a consistently unsuccessful prediction over the last few decades, so why was I following this bad precedent? But it looks like I timed it more or less right. My desktop and laptop are both 2 years old. One of the corollaries of Moore's law is the expectation that performance will double about every 2 years, so I should be able to go out and buy new machines that are twice as fast.

Sadly, I can't. Now I know that clock speeds are a notoriously bad way of predicting performance, but increases in clock speed tend to overpredict performance, i.e. a doubling in clock speed typically delivers less than a doubling in performance, all other things being equal. So given that my desktop is a hyperthreaded 3.04GHz Pentium 4, it's probably a safe bet that a 3.73GHz hyperthreaded processor, one of the two candidates for 'fastest desktop CPU' on offer from Intel isn't going to be twice as fast. The other candidate is a 3.2GHz dual core unit. I guess the theoretical maximum throughput of that CPU might be double that of a 3.04GHz single-core hyperthreaded unit, but not in a way I'm likely to use on my desktop - I rarely see 100% CPU load on my hyperthreaded CPU because I rarely have 2 threads running simultaneously, so for the workloads I subject my desktop to today, dual-core doesn't help me.

(Yes, dual-core does mean that strictly speaking Moore's law, which said nothing about clock speeds, is still going. But it's the performance corollary of Moore's law that most people care about.)

It's the same story on the laptop front. I have a 1.6GHz Pentium-M. And yes you can buy Pentium 4M CPUs with clock speeds that are over double that, but the M and 4M performance-to-clock-speed ratios are very different. You get about the same amount of grunt from a Pentium M as you get from a Pentium 4M with between 1.5 and 1.75 times the clock speed, depending on what exactly you want to do with it. The most meaningful comparison is to compare one Pentium M with another, and the fastest Pentium M money can buy is clocked at 2.13GHz, which is only about 33% higher than my current 2 year old machine.

And yes there are other advances - faster FSBs, and probably some tweaks to the architecture. But these aren't going to get me to a doubling in performance.

So it has begun. Some people are in denial - you regularly hear people invoking Moore's law when justifying some inefficient design. But many are noticing - Herb Sutter's seminal article seems to be the definitive essay on this. Figure 1 in that article is particularly striking - the line showing the ceiling on clock speeds exaggerates the kink somewhat (notice that the underlying points don't draw quite such a clear picture as the best fit lines) but still illustrates that progress in the last few years has failed to keep up with the expectations Moore's law has set for so long.