Underclocking, going backward?

Do you remember back in the days when a Pentium III doing 600GHz was awesome? At that time, when guys at Intel were foretelling that the increase of the processors clock rate will have no end, or at least none that they could possibly envision, you’d see that oath as testimony of the faith in a future of endless possibilities, gaming-wise.

Weirdos...

Later on, Intel went as far as publishing a Pentium 4 which was degraded version of the Pentium III. Less complex, less instructions, it was able to go higher in clock rate than any Pentium III, something like 1.8GHz easily. It went on. People even bought laptops with Pentium 4 2.6GHz. And then people started noticing: hey! it’s winter, it’s freezing damn cold outside but I’m not even forced to turn the heat on! Or funnier, hey, why do my brand new laptop is making more noise than a vacuum cleaner? And what black magic made power supplies became a noticeably costly component of a computer?

Well, that’s all about physics. And there’s not much to do about. The faster the computer processing unit run, the more energy it will burn, the hotter it will get.

AMD was smart enough to soon start shipping processors with lower clock rate than Intel ones for the same effective potential. It was also smart enough to brand them accordlingly, branding them for instance something like 3200+ to tell they would be as potent as a Pentium 4 3.2Ghz, while they had a way slower clock rate.

Intel could surely not completely obviously go backward – and publicly recognize AMD wittier. But they could not loose the growing market of the laptops, where the heat issue (not to mention the impact on the batteries life) was too much of a problem with Pentium 4, so they invented the Pentium M… based on the Pentium III, of course.

Considering the unavoidable antagonism beetween fastness and energy consumption, the best idea that someone (who, I do not know) came up with was to enable the operating system to set the clock rate according to the current need. It comes with many different names (Cool & Quiet, whatever) and I believe is it now available with most recent processors. On Debian, you just have to install cpufredutils and load the relevant kernel module cpu (powernow_k8 for instance on my AMD Athlon 64 X2 Dual Core workstation) and then pick a policy. Yes, you have to pick a policy, like on demand, performance, etc.
Obviously, there is a performance loss (hence the name –performance– of the policy which actually only set the clock rate to max) since you are not always running the fastest possible: there is always a delay needed for the operating system to understand that now you need full speed when it was idle just before. The different policies purpose is to optimize this delay – tuning inertia, in which regard on demand is simply harsher than conservative.
Next step would be to have the operating system guessing if you’ll need full speed or not according to what you are actually doing (which software do you run, etc) and what you are about to do according to past usage (yes, logging what you use to do and making guess).

So currently, on a workstation like mine, using cpufreq on demand is probably a wise choice. Most of the time, it will run slower than it could, because you do not need full power of a recent processor to browse over the web, reply to mails and whatever crap like that you may want to do. And when you’re compiling a piece of code, when you are encoding a piece of music, then you’ll have full power. I never or rarely use GNU/Linux to play games so inertia is not a crucial issue – however, to play games, it would surely be best to set the policy to performance, even if after the game is started it will likely, anyway, request full power (surely, you configure your games to the best resolution, anti-aliasing, etc, that your box can handle, don’t you?).

(Not to mention that, gaming-wise, graphical cards now do a big part of the job, the most important anyway, making CPU less important by comparison to so-called GPU… but that’s another story)

So, now, I’ll get straight to the point. I run also a little shuttle box as local server. It serves files, it is up 24h/24 and do plenty of small things. It comes with a Celeron 2.6GHz but it surely would do as well with a slower clock rate. With in mind the idea of reducing the heat of this processor as much possible, I searched over the web on the subject of underclocking. The mainboard of the shuttle, by design, does not allow to make this processor run slower than it does. There is no possibility of playing with cpufreq or alike with a Celeron – which is actually a crappy Pentium (no L2 cache, less instructions, etc).

Pentium 4 2.80GHz running at 1.40GHz

I found however interesting the idea of buying a processor designed to run with a faster front side bus than the actual mainboard we have. It focus on the fact that the processor clock rate is actually determined by both the clock multiplier and the front side bus (FSB).
This shuttle front side bus runs at 400MHz. If I pick a processor, says, designed for a 800MHz front side bus, which is usual of Pentium 4 around 3GHz, it will run twice slower.

So I spent nine euros to get a (used – but the Celeron 2.6Ghz is not brand new either) Pentium 4 2.8GHz. And now, my shuttle runs 1.4GHz. Processor temperature is around 35°C, and the sole fan of the box is around 1300RPM. Nice side effect, this processor got Intel’s Hyper-Threading (simili multiprocessor), which is definitely good for a server.

I guess I could even use cpufreq with this new processor. But I’m not sure it would let me go as far as reducing by a factor of two the clock rate. And anyway, I’m happy in this case with such clock rate set up by hardware.

The only remaining thing to do is to undervolt it now.