A little undervolting adventure

[Queelis]

Hello there,

there are a lot of posts I've seen where people say "I've undervolted my CPU" and all, and about how Phenom IIs undervolt well, but no threads which would give some detail about what to expect, so decided to do it myself. After all, how hard can that be (and what could possibly go wrong)?

By the way: couldn't figure out which forum to put this to, so just took my best shot

So, starting out. The hardware is on the left if you like simple lists, but I'll post everything here as well, should anything change. I have an AMD Phenom II X3 720 BE in a Gigabyte MA790X-UD3P motherboard with 4 gigs of Corsair RAM, a Sapphire Radeon HD 4830, everything seated cozily in an Antec Sonata Plus. The cooling is stock, except that I've swapped the Antec Tricool case fan for a Nexus 120mm fan rotating at some 1000 RPM. The OS is Vista 64bit.

Since I have no means of measuring the power consumption of my PC, I've resorted to just checking the temperatures of the CPU as means of seeing the (supposed?) "benefit" of undervolting.

The drill was to load the CPU using OCCT's Linpack test for 2 hours - since idle temperatures would be very similar anyway, only load temps will have to do. And I felt 2 hours would be enough to get the highest temperature the CPU would reach at a selected voltage. For some reason, the 0.025V less voltage test was done only for an hour. Anyway, here we go.

First run, stock voltage (which is 1.325V, by the way)

Peaking out at 46-47 degrees on Linpack - that is pretty cool (no pun), considering I only have stock cooling. The CPU fan spins at 3200RPM, though - pretty harsh on the ear, ouch.

Second run, at the smallest possible increment - -0.025, which leaves us at 1.3V.

One increment, approximately one degree drop in temperature - 44-45 degrees Celsius.

Carry on, -0.050 (1.275V)

Again, a degree shaven off (43-44C)

-0.075 (leaves 1.25V)

Hovering around 41-42 degrees. Quite good.

-0.100 (1.225V)

40-41 degrees. Still decreasing a degree every increment.

-0.125 (1.2V)

39-40 degrees! This is turning out quite well.

-0.150 (1.175V)

Well, it had to end some time. After some 12 minutes into the test, an error is reported and the test was stopped.

So, that left me with a lowest possible voltage of 1.2V. To be sure I had it stable, I set OCCT custom Linpack test to run for eight hours, and it completed it without any errors:

The temperatures don't even break the 40 degree mark during the 8 hour run - I'm happy with that!

Since all these test were run on different days, I wanted as accurate a comparison as simply possible, so I've re-run the stock and 1.2V setting tests on the same day, for the sake of completeness. The results are:

Stock: 46-48 degrees, 1.2V - 40-41 degrees. So that's 6 degrees lowered temperatures with nothing just some free time on one's hands

So there you have it, I had some safe fun fiddling with my PC (I'm happy with the performance so didn't want to overclock), the temperature drops are nice. And the fan now spins a bit slower, which is actually audible (3200RPM v. 2800RPM). So, cool and quiet(-er, -ish. Whatever), just the way I like 'em Currently I'm at 1.2V load voltage, and haven't had issues (maybe anything I do just isn't as harsh as Linpack).

However, I've had some issues while doing all this testing, and I've started a new thread so as not to clutter everything into one, hope you'll head there to see if you have some solutions!

[lego batman]

I suppose the next thing to try would be underclocking it as well! Might give you some more room downwards until you find the minimum that the chip can run at all.
Good job there! That's some dedicated methodical research.

[Queelis]

Mmyes, well, I could. But then again, my goal was to get the lowest possible voltage without sacrificing performance, since I could just leave the CPU at x4 multiplier and have 0.8-0.9V at load, but that wouldn't be much faster than my previous Pentium III

[Queelis]

Long-time usage update: at 1.2V, there were some times that I would get artifacts and strange music playback issues, Firefox window getting black bars and strange things like that. But that was once-twice in those two months. Now I've recently started folding@home (with GPU and CPU), and seeing that f@h SMP client puts a great load on the CPU, it threw an error when I was 80% into a work unit

So, new stable (I hope) voltage: 1.225V! That's still a whole .1V less than stock, though. The load temperature with both the GPU and the SMP clients running is 41-42C on the stock heatsink, which is too loud for my liking, but temps are okay.

[Queelis]

Hello again,

thanks to some luck, I've had the chance to get a Phenom II X4 965 BE for an upgrade from previous X3 720 BE (though fully unlocked, but nevertheless).

That said, it's a 125W unit as opposed to my 95W 720, so wanted to undervolt that as well.

As a reference, I've registered that my unlocked X3 720 would happily run four cores at 2.8GHz with 1.275V (down 0.05V from stock voltage of 1.325V) with a temperature of 46C with the fan running at 1600RPM.


Sadly, my stock data of the new CPU is unreliable, since it was gathered with EasyEnergySaver enabled, which resulted in less voltage and as such lower temps and fan speed RPM, but I'll include it nonetheless.

So, here we go:

Stock voltage, 1.35V:
Average temperature: 51.7C
Average RPM: 1974

-0.025V, 1.325V:
Average temperature: 50.6C
Average RPM: 2287

-0.05V, 1.3V:
Average temperature: 50.3C
Average RPM: 2306

-0.075V, 1.275V:
Average temperature: 49.5C
Average RPM: 2277

-0.1V, 1.25V:
Average temperature: 48.5C
Average RPM: 2243

-0.125V, 1.225V:
Crash

So here we go, again the CPU runs along smoothly (at least for now, it's been on load for several hours, so probably will be OK) with 0.1V less than intended, which lowers the temperature by around 4 degrees. Currently I've re-enabled EES and am currently running the fan at 1900RPM (completely reasonable noise) with the temperature hovering around 48-49C.

Interesting thing about the difference a stepping/revision makes: at the moment I'm running the CPU at 600MHz more than the previous one with 0.025V less running through it. Win-win

So much for now.

Cheers,
Queelis


EDIT: nvm, glitch.

[Ulti]

As my board didn't like SpeedStep and I didn't need 4ghz for what I do (Well technically an i7 920 and 5850 is overkill for me haha) I decided to do something similar a few weeks back too.

I undervolted my i7 930 but at the same time pushing it as far as it could. This resulted in 1.000~1.023V as the lowest voltage and a speed of 3.2Ghz. I am prety happy with this result! Load temps just come under 60 degrees which is very good for an i7 930 imo.

[Queelis]

Wow, that's incredibly low. What's the stock voltage of the i7 930?

[Ulti]

Stock voltage tends to be around 1.200V-1.265V or something AFAIK, not sure what the VID of this particular chip is though.

[Attila the Bun]

Queelis,
Just a quick question from someone completely devoid of underclocking knowledge. What is the object of underclocking? Is it merely an exercise in energy efficiency? Or has it further advantages that aren't obvious to me?

[barker967]

Less volts means less heat produced as the cpu is using less power to function, obviously if you go below a certain voltage point the cpu wont be able to function at all as it doesn`t have enough power. Opposite when you are overclocking, you need more volts as you are pushing the cpu to perform over it`s intended power point, downside of that is you produce a lot more heat thus better cooling is required.

[scaryjim]

Should probably also point out that Queelis hasn't been underclocking, just undervolting. Basically as the physics goes, reducing voltage also reduces current drawn and therefore exponentially reduces the power draw. A side effect of reducing the power draw of the CPU is that it also reduces the heat dissipated by the CPU, which in turn should help the longevity of the CPU ('cause heat + electronics is bad, m'kay ).

It is possible to also underclock a CPU to reduce power draw, but this is a linear reduction (incidentally, both AMD's Cool'n'Quiet and Intel's SpeedStep technologies both automatically underclock *and* undervolt your CPU at idle to save power ), and obviously reduces performance, so the point for Queelis here is to find the lowest voltage at which the CPU operates at stock speed: reducing both power draw and temperatures while maintaining performance.

[Attila the Bun]

Should probably also point out that Queelis hasn't been underclocking, just undervolting. Basically as the physics goes, reducing voltage also reduces current drawn and therefore exponentially reduces the power draw. A side effect of reducing the power draw of the CPU is that it also reduces the heat dissipated by the CPU, which in turn should help the longevity of the CPU ('cause heat + electronics is bad, m'kay ).

It is possible to also underclock a CPU to reduce power draw, but this is a linear reduction (incidentally, both AMD's Cool'n'Quiet and Intel's SpeedStep technologies both automatically underclock *and* undervolt your CPU at idle to save power ), and obviously reduces performance, so the point for Queelis here is to find the lowest voltage at which the CPU operates at stock speed: reducing both power draw and temperatures while maintaining performance.

I suppose you want another thanks now don't you.

Actually nice explanation...does this process have the same sort of fallout as overclocking ie. system instability and cpu death or is it more forgiving than that?

edit: AHha...yes it has undervolting in the title.

[barker967]

System instability, very unlikely to damage a chip running too few volts through as it`s high voltages that will burn the cpu not low. Random bsod`s and prime95 errors.

[Terbinator]

So while not under-volting in the same sense, I'm running my GPU (see system on the left) @ 0.975v (down from 1.025)while maintaining the factory O/C

Will report back when i see how low it can go.

[scaryjim]

Yeah, you can definitely end up with system instability when you undervolt. I have a Sempron 140, which is one of those lovely AMD processors made by disabling bits of a bigger processor (in this case it's a locked Athlon II X2). Mine unlocks happily to a dual core at stock voltages, but at reduced voltages I get instability while doing certain CPU-intensive tasks (particularly while watching non-GPU-accelerated flash video). As a result I now run it locked as a single core but at reduced voltage (down to 1.025 from 1.325) for some fairly significant power savings (about 10% at full load for the whole PC!).

[Queelis]

Should probably also point out that Queelis hasn't been underclocking, just undervolting. Basically as the physics goes, reducing voltage also reduces current drawn and therefore exponentially reduces the power draw. A side effect of reducing the power draw of the CPU is that it also reduces the heat dissipated by the CPU, which in turn should help the longevity of the CPU ('cause heat + electronics is bad, m'kay ).

It is possible to also underclock a CPU to reduce power draw, but this is a linear reduction (incidentally, both AMD's Cool'n'Quiet and Intel's SpeedStep technologies both automatically underclock *and* undervolt your CPU at idle to save power ), and obviously reduces performance, so the point for Queelis here is to find the lowest voltage at which the CPU operates at stock speed: reducing both power draw and temperatures while maintaining performance.

Spot on I'm satisfied with what I have and don't need more power (well, more folding@home points would be nice, but I don't actually need them ), so I might just as well get that computing power as efficiently as possible. But I don't want to underclock, since then there would have been no point in upgrading (from a 720BE).

And yes, reduced temperatures means reduced fan speed, i.e. noise.

It's as simple as reducing a voltage, running a stress test (half an hour is OK for general understanding that you can go lower), lowering, testing. When the PC crashes (usually in a matter of minutes under stress), on reboot kick the voltage back up a notch, run stress test for a while to ensure the stability (my 965BE has been folding for days, for example, and hasn't crashed once).

No reason not to

So while not under-volting in the same sense, I'm running my GPU (see system on the left) @ 0.975v (down from 1.025)while maintaining the factory O/C

Will report back when i see how low it can go.

Why isn't it the same? Having the same performance for less power, heat and noise - that's exactly what I'm doing