Huge success undervolting the Air!

I read the thread about down-clocking to 1.2ghz under thermal load, installed a program to monitor the *actual* core speed, and saw exactly what was going on.

I think this is a software tuning problem. Apple was in a rush to get the Air out and decided to err on the side of stability rather than performance. This is probably the right decision, but it doesn't mean that us hackers can't do a little custom tuning ourselves to get the Air to perform a little better.

So I installed CoolBook:

http://www.coolbook.se/CoolBook.html

It costs $10 for the version that lets you tune, so I coughed it up and started tuning.

First, if you don't know what undervolting is, it's basically the inverse of overclocking. When you overclock you are boosting the clock speed (and often the voltage) to pump up your chip closer to its *real* performance cutoff, often using bigger heatsinks and such to dissipate the extra heat. When you undervolt you're going for the bottom end of your chip's voltage/speed stability curve. Undervolting does *not* slow down the chip, but can make it unstable if you go too low just like overclocking can make it unstable if you go too high.

CoolBook allows you to tune the processor's voltage and speed throttling settings. Be careful about heat here, since it also allows you to ease back on the thermal throttling. (Of course, I think at 100C the chip will do emergency core shutdowns so you're probably fairly safe.)

CoolBook is nice, but it has an absolutely awful user interface. The "add" button also updates existing entries, and if you delete the 800mhz by accident you have to hit "default" to restore the factory default settings to get it back.

So it took me a few hours to get it right. Each chip may be different, so your mileage may vary. Here are my settings:

On adaptor:

1200MHz: 0.9750V
1400MHz: 0.9875V
1600MHz: 0.9875V

On battery:

800MHz: 0.9000V
1200MHz: 0.9750V
1400MHz: 0.9875V
1600MHz: 0.9875V

Thermal limit: 85C (the max)
Throttling level: high

(The throttling level seems to be how often the system checks and re-adjusts the frequency. Higher is more often.)

The factory default for 1.6ghz is 1.15V, so this is substantially lower... and boy does it make a difference! Also notice that I deleted the 800mhz entry for adaptor mode. I found that this made the machine a little snapper, and this entry is for battery saving. I left the entry in for battery.

(If you wanted you could remove the higher speeds in battery mode with CoolBook and probably get better battery life at the expense of speed while on battery. If I were on a long flight I might remove everything but 800mhz, especially if I was just going to be typing. 800mhz is fast enough to watch movies, word process, code, and pretty much most other routine stuff.)

If you don't know what these entries mean, this is the chip's throttling curve. When it's not heavily loaded, it throttles down to the bottom to save power and heat. When it is heavily loaded, it throttles up to max unless the thermal limit is tripped. In this case, it throttles down a notch until the temp is under control.

At these voltages, I get 1.6ghz consistently when on internal monitor only and as long as I don't block the rear vent for too long. With an external monitor connected, it bumps the limit and throttles down but only after many minutes of sustained maximum CPU load... and not for long. It'll bump down to 1.2 for a few seconds and then back up to 1.4 and 1.6 to hover around 85C. I bet if I were to better elevate the machine I could get it to sustain 1.6ghz continuously with a monitor connected.

It runs a *lot* cooler!

CoolBook will let you turn thermal limit off, and this allowed 1.6ghz to be sustained forever with an external monitor (of course). However, seeing the chip reach 92C makes me nervous so I turned it back to 85C. (100C is the chip's emergency shutdown level, but 90+ is uncomfortably hot for my taste. I wish CoolBook had a 90C option though.)

Getting these settings required some trial and error. First I set everything to 0.9V, but I got instability. So I edged the curve up until my machine seemed stable even when doing fourier transforms and other abusive things.

Be aware: the instability you get with undervolting (and overclocking as veterans know) can be weird. It can include: application crashes, funky screen corruption, slowdowns / system becoming un-responsive, and kernel panics of course. If you experience any of these symptoms where you haven't before, edge up your voltages a bit.

So you brave souls can give this a try. I bet it'll improve battery life too.

Epilogue: since I fixed it with software, Apple could probably fix the Air's heat/performance issues with an update. However, doing this would be *very* risky as the undervolting ability of chips varies. Like I said, Apple clearly chose to err on the side of stability. This is a good choice for them, since most users of the Air aren't performance-hungry but would be very irked about crashes. But it does seem to me that they could get away with tuning this a *little bit* (less agressively than me) and probably make the egregious things like core shutdowns go away.

atszyman said:

I don't think it has anything to do with rushing or erring or decisions. The processors themselves work over a wide range of voltages and speeds. The voltage and speeds are designed to hit certain targets that the most chips will run at. Many chips will go faster, many will operate just fine at lower voltages. However in order to ensure proper operation of the chip, for every chip they install, they have to meet the criteria set forth by Intel. Did you have good luck with yours? Yes, although other people's mileage may vary and they will be unstable at your settings, or some might even find they can go lower voltage than you and still remain stable.

Click to expand...

That's exactly what I meant, but yeah. YMMV

I wonder if Apple could test the chips in the factory and custom-tune each one? That would probably add an expensive/long step though and raise the manufacturing price.

Depending on the voltage your adjusting you could fry the chip by under-volting. Most chips these days require a few different voltages to run, and if you lower one, there are sometimes sneak paths where the logic that is not getting enough power can draw power from the I/O voltage, or other auxiliary voltages on the chip which can burn out IO and other parts since they were never designed to carry the amount of current drawn through the sneak path. There are usually some protections in place to prevent the burn out, but with millions of transistors on a chip it's sometimes hard to catch all of the possible sneak paths.

Click to expand...

Has anyone ever heard of undervolting actually damaging a chip? I've done some Googling on the undervolting subject and I've never heard a fry story, but of course this could make the chip unstable.

I started a thread on the damage issue over here:

http://reddit.com/info/6dbq4/comments

I'm not saying that it has happened with processors, but with any IC/sometimes even circuit boards, you can run into issues with "sneak paths" that try to power up chips through the IO pins that will most often cause erratic operation, but have the potential to damage a chip if the proper protections are not in place. Usually you'll see this when a chip that drives an input to another chip that has not yet received power. The non-powered chip will try to draw it's power through the driver pin's output. Most chips these days have protections in place to avoid burning out the IO pins from drawing too much current through the pad, but there are other ways that these "sneak paths" can occur. I've seen this happen in my company's products. Many times, if you read the data sheets, parts will have very specific power sequencing guidelines to ensure that the core voltage comes up before the IO to avoid potential problems of this nature.

I'm not saying that under-volting will damage the chip, Intel may have all of the bases covered with adequate protections in place, but depending on the chip and how it's set-up the power sequencing and under-voltage can cause erratic operation or damage in some circumstances. I'm just supplying the information so that no one thinks this is completely risk free. Chances of damage are less than over-clocking and over-volting, but there is a non-zero probability that damage can occur.

Click to expand...

That along with some of the comments over in Reddit have clarified my understanding a bit. (I'm not EE, but I do know a bit about electronics and a bit more about physics.)

Obviously there are *no warranties* here. YMMV! But, I consider it highly unlikely that a chip like the Air's Core 2 Duo would experience what you and some of the other EE folks I've asked are describing. The chip in the Air is designed not only to run across a range of voltages and speeds but to switch voltages and core speeds on the fly-- which it does constantly. None of the voltages CoolBook will let you switch to are outside the spec'd range, and a chip that switches this much is probably designed for a bit of voltage jostling.

But you're right to point out that from a physics/EE point of view there is a nonzero probability of damage... but I think it's probably quite low and I couldn't find an incident report of undervolting damage with Google. (Overclockers who overvolt on the other hand have burned chip galleries and they brag about it like car people brag about blowing up engines... but overvolting is a whole 'nother matter due to heat!)

atszyman said:

I've seen PCI bridge chips completely fried due to a power sequencing issue, so I just wanted people to be aware that this is not a risk free proposition.

Click to expand...

... you're right, and thanks for the warning.

But one more point I thought of: CoolBook's tuning is done by a kernel extension and a daemon called CoolBookSwitcher. These start after the mach kernel and the core extensions boot, so when the machine is turned on it's running it's factory default settings until then.