Max Undervolt - Phenom II X4 965 BE (C3)

[jvroig]

Using the stock AMD Cooler (looks ok, but fan is noisy and annoying when it reaches ~5000 rpm), I undervolted it to 1.232V (1.250V in bios) from stock voltage of 1.360V (1.400V in bios). I use CPU-Z and the EasyTune (Gigabyte utility) to read the vcore in Windows. In Linux, the lm-sensors and gnome sensors applet report the same vcore as CPU-Z and EasyTune.

1.232V vcore is Prime95 stable for over 12 hours, and been using it for two weeks and had no troubles.

1.216V vcore in CPU-Z/EasyTune = Didn't pass Prime95 for more than 6 hours. No errors, but when it reaches around 6 hours, it just reboots. I tried this 2 times to be sure.

1.186V - still boots ok. However, this time it doesn't stand up to Prime95 for even an hour. I watched it twice, and when it starts running Prime95, less than twenty minutes later it will reboot. No errors, no bluescreen, nothing. Just reboots.

So I went back to the lowest stable vcore (1.232V). There were no difference in temps anyway (1.232V vs 1.216V), and even against 1.186V, there was at most only 1C advantage for 1.186V, but seeing as to it can't even do Prime95 for an hour, I'm not exactly crying about it.

I'm more curious about 1.216V though. Since it does manage to do Prime95 for a few hours, realistically speaking I guess this is a safe undervolt? There is no real-world usage for me that loads all cores at 100% steady for even an hour. Depending on what I'm doing, I usually get 1-2 cores at 50-80%, while cores 3 and 4 are at 10-20% (according to gnome system monitor, which I have on all the time in a separate workspace), but that's about it. And when a core does get 100% utilized (report generation is usually the only thing I do that does that), it's really just a single core that gets loaded much, the rest are still in their usual 10-30% loading. So I'm thinking even 1.216 in my "real-world" usage would feel just fine.

It's just my thinking about it, though, and wanting to hear your thoughts as well. This is the first time I even spent time undervolting, to manage temps better.

Using the stock cooler and at 1.232 vcore, and with high ambient temps (around 86-88F or 30 - 31C; I'm in a hot tropical country), CPU idles at 39C, and at max load (Prime95 for several hours) temps max out at 59C. Yep, stock cooler. I'm impressed. Makes me wonder why AMD had their X4's at such a high stock vcore when they can live on much lower voltage and enjoy better temps.

Anybody else here undervolted their processors for good measure?

Also, I've read lots of conflicting information about undervolting being safe or as dangerous as overvolting. Some claim it is definitely safe, and the only worry is stability. Some claim it is as dangerous as overvolting and overclocking, since you are still doing the same thing in principle: "maximizing the MHz while minimizing the voltage". I've even read one of the proponents of undervolting being as dangerous as overvolting say that as you lower the voltage, you end up increasing the amperage.

I tend to believe undervolting is far from being as dangerous as overvolting and overclocking, but don't let that stop you from telling me otherwise as I am no expert at all.

 

[alyarb]

the dangers of overvolting are the localized effects of heating due to increased leakage current and the displacement of interconnect material due to increased current density and current crowding. undervolting implies less current and less heat. it does not stress the material in any way. trying to start up a CPU at 4 ghz without enough power doesn't hurt it. it just isn't going to run.

running the same CPU with a little* too much power doesn't "hurt" it either, but you sustain a greater rate of occurrence of these phenomena over time and the lifespan of the circuit is exponentially reduced. interconnects made of copper are good conductors, but they do resist, because even the purest copper wire will have a nonuniform lattice structure for an arbitrary length. this is why there is electrical resistivity in metals. charge carriers flowing along this nonuniform lattice impart nonuniform amounts of momentum to the lattice elements (copper atoms). the atoms therefore incur nonuniform vibration relative to one another and this creates heat. sometimes the atoms can be moved a little bit, and if this happens enough and an important junction is ruined, your CPU could be permanently unstable or (more likely) will never run again. as these interconnects go from the 90 to 32 nanometer scale, the diameter of the interconnect becomes incredibly fine; so not only do these thin conductors resist, but they must resist more current-per-area than their siblings with larger transistors. therefore at a given power level, you can see how current densities increase stupendously (in high power CPUs with billions of transistors, the tiny mass of an individual electron begins to really, really add up at these densities), and stupendous operating conditions will kill a CPU.

This is one of the reasons why CPUs of all kinds will never see super high frequencies. This is a natural law and we just have to make the best of it. Not only that, as transistors become vanishingly thin, you reach a point where almost any conventional current density used in previous CPUs is too high. If you think we hit a wall in 2006, just wait. There are many more walls to hit. To get out of this hole, we will either need a superconducting interconnect that works at room temperatures or higher (never going to happen unless there is literally some miracle with carbon nanotubes) that can be deposited with the utmost precision (an entirely separate miracle)... or use photons as our charge carriers, which is more ambitious but also more realistic and nevertheless requires a completely clean slate in terms of design, materials, manufacturing techniques, etc. i'm sure more progress has been made than i'm giving credit for because i just don't read about this stuff, but there are people somewhere, like intel, ibm, and universities for chem dorks that have been exploring ways out of this shithole for decades.

*never set your CPU to an arbitrarily high voltage (such as 1.896) "just to see how hot it gets." 1.9 volts is too high. even 1.6 volts is too high. you're probably lucky it crashed because if it ran for hours and hours it probably would've damaged it without sub-zero cooling.

 

[jvroig]

Oops, sorry. That was a typo. I was undervolting... I meant ~1.186V, the next step below 1.216V that my mobo allowed. I will make an edit in my OP so as not to confuse any other reader.

Thank you for the very well-explained post. The dangers of overvolting are not alien to me, but your post did so much to enlighten me as to why.

And I take it that you also said, definitively, that undervolting in no way damages or reduces the lifespan of the CPU, yes?

Thanks.

 

[soccerballtux]

the dangers of overvolting are the localized effects of heating due to increased leakage current and the displacement of interconnect material due to increased current density and current crowding. undervolting implies less current and less heat. it does not stress the material in any way. trying to start up a CPU at 4 ghz without enough power doesn't hurt it. it just isn't going to run.

running the same CPU with a little* too much power doesn't "hurt" it either, but you sustain a greater rate of occurrence of these phenomena over time and the lifespan of the circuit is exponentially reduced. interconnects made of copper are good conductors, but they do resist, because even the purest copper wire will have a nonuniform lattice structure for an arbitrary length. this is why there is electrical resistivity in metals. charge carriers flowing along this nonuniform lattice impart nonuniform amounts of momentum to the lattice elements (copper atoms). the atoms therefore incur nonuniform vibration relative to one another and this creates heat. sometimes the atoms can be moved a little bit, and if this happens enough and an important junction is ruined, your CPU could be permanently unstable or (more likely) will never run again. as these interconnects go from the 90 to 32 nanometer scale, the diameter of the interconnect becomes incredibly fine; so not only do these thin conductors resist, but they must resist more current-per-area than their siblings with larger transistors. therefore at a given power level, you can see how current densities increase stupendously (in high power CPUs with billions of transistors, the tiny mass of an individual electron begins to really, really add up at these densities), and stupendous operating conditions will kill a CPU.

This is one of the reasons why CPUs of all kinds will never see super high frequencies. This is a natural law and we just have to make the best of it. Not only that, as transistors become vanishingly thin, you reach a point where almost any conventional current density used in previous CPUs is too high. If you think we hit a wall in 2006, just wait. There are many more walls to hit. To get out of this hole, we will either need a superconducting interconnect that works at room temperatures or higher (never going to happen unless there is literally some miracle with carbon nanotubes) that can be deposited with the utmost precision (an entirely separate miracle)... or use photons as our charge carriers, which is more ambitious but also more realistic and nevertheless requires a completely clean slate in terms of design, materials, manufacturing techniques, etc. i'm sure more progress has been made than i'm giving credit for because i just don't read about this stuff, but there are people somewhere, like intel, ibm, and universities for chem dorks that have been exploring ways out of this shithole for decades.

*never set your CPU to an arbitrarily high voltage (such as 1.896) "just to see how hot it gets." 1.9 volts is too high. even 1.6 volts is too high. you're probably lucky it crashed because if it ran for hours and hours it probably would've damaged it without sub-zero cooling.

electromigration has hardly ever been experienced in real life though...for the majority of users, in fact, who do not use their processor hard for more than a few hours/day, their processor is going to last, with overclocking, _longer_ than a server CPU that hits 100% or some other processor that's constantly pegged-- because the electromigration is a function of current consumption. In other words, it's not to be worried about, in my opinion.

 

[heyheybooboo]

.....

Anybody else here undervolted their processors for good measure?

.....

I tend to believe undervolting is far from being as dangerous as overvolting and overclocking, but don't let that stop you from telling me otherwise as I am no expert at all.

I think Phenom IIs tend to be a bit over-volted at stock. As long as your under-volted proc passes whatever stability testing you are comfortable with I'd say you are GTG.

I've got an unlocked, under-volted Phenom 550 in the 1.2v range. Power consumption at load is 1/3 less (127w) than stock ---- with a Gigabyte 785g (which I think kinda over-volts, anyway).





--

 

[frostedflakes]

Yeah, the Gigabyte boards actually overvolt the Phenom II. Stock voltage is 1.325V, but they run them at 1.4V.

Anyway, been undervolting for years, it's a great way to reduce heat and power consumption without impacting performance. And with underclocking and undervolting, you can really make your CPU sip power, but of course the tradeoff here is performance.

I have a Phenom II X2 550 BE that I'm running at 3.1GHz and 1.175V. I've also played around with unlocking the cores, and found that running four cores at 2.6GHz and 1.025V uses the same power from the wall as the dual core config at 3.1GHz and 1.175V.

 

[joetekubi]

Thanks for the discussion, guys.
For my upcoming Phenom X4 build, power consumption, longevity, heat, and noise are much more important to me than Ghz numbers. I'll try undervolting and post results here after the build.
--Joe

 

[alyarb]

electromigration has hardly ever been experienced in real life though...for the majority of users, in fact, who do not use their processor hard for more than a few hours/day, their processor is going to last, with overclocking, _longer_ than a server CPU that hits 100% or some other processor that's constantly pegged-- because the electromigration is a function of current consumption. In other words, it's not to be worried about, in my opinion.

a highly overvolted CPU that sees a lot of idle time doesn't have to worry about long-term effects of current crowding, but it is a continuous function of current consumption and as time goes on the effects become less subtle, just not enough to break an interconnect. a lot of SETI dorks were the first to document the sudden northwood death syndrome, but they were doing 3.2+ ghz at 1.75-1.85 volts. it doesn't depend so much on the "size" of the CMOS node that let them down as it does on the current density relative to the given geometry of whatever the weakest junctions were, but 1.8v sounds too high to me for that CPU. server/HPC CPUs that are pegged their entire life contend with this too, but as long as they are running @ stock the rate of degradation will be quite slow. it would be interesting to get an electron microscope video of this in action just to see how long this process can take.

We regular people are safe from electromigration for other reasons, like people who are into distributed computing for the fun of it are compulsive about running the lowest voltage they can find simply for conserving energy, and most overclockers stay reasonably within the envelope suggested by other specimens from the same CPU family, and like you say, usage patterns that do not need large currents for long periods.

 

[heyheybooboo]

Yeah, the Gigabyte boards actually overvolt the Phenom II. Stock voltage is 1.325V, but they run them at 1.4V.

Anyway, been undervolting for years, it's a great way to reduce heat and power consumption without impacting performance. And with underclocking and undervolting, you can really make your CPU sip power, but of course the tradeoff here is performance.

I have a Phenom II X2 550 BE that I'm running at 3.1GHz and 1.175V. I've also played around with unlocking the cores, and found that running four cores at 2.6GHz and 1.025V uses the same power from the wall as the dual core config at 3.1GHz and 1.175V.

Good job!

I tested a bit around 1.18v and it worked great. The issue I had was at the low end (something around 0.84v-0.86v).

Most of the time it would work great, but 20% of the time it would blue screen coming out of sleep. I don't really think it was a specific issue with the proc/voltage but trouble with the Gigabyte mobo and how it handles sleep/hibernation.

I flashed to the latest BIOS (F7?) but it seemed to make the problem worse.





--

 

[frostedflakes]

You know I thought I tested 1.175V shortly after I got my processor and had stability issues (occasional BSODs and restarts, stuff like that), so I had to bump it up to 1.2V. But then like a month or two ago I gave 1.175V another try and haven't had any issues yet, knock on wood. Maybe the colder winter temps helped. I think you're right that the issue was with the lower CnQ power states, because even before I could Prime95 all day at 1.175V. The only time I would have restarts was during light/normal use. I don't know if I had that sleep issue, though, I've never tested that function.

I'd have to double check, but I think I'm using BIOS F5. Motherboard is Gigabyte GA-MA785GM-US2H, by the way.

 

[jvroig]

Motherboard is Gigabyte GA-MA785GM-US2H, by the way.

That's mine too. Got it so I can re-use all of my 4GB RAM from my old system (Athlon X2 7750, MSI board) to the Phenom II X4 system, saves a lot of cash now that RAM prices don't seem to want to go back down to their prices last year.

For my upcoming Phenom X4 build, power consumption, longevity, heat, and noise are much more important to me than Ghz numbers.

For the first time in a build, noise became a problem as the Phenom II X4 stock cooler was quite unlike the ones in the old Athlon X2 5000 and even the 7750. It whirs like crazy at full speed (6750 RPM), and even at 5000RPM it starts to be very noticeable and annoying. I was going to get a more decent cooler, but my options were limited thanks to my case, and I was thinking twice about it because I wasn't going to overclock. So I tried undervolting to help manage temps, and I was very satisfied with the result. Despite high ambient temps and the stock cooler, the processor is chilly at 39C idle and no more than 58-59C as the absolute max temp when running Prime95 for several hours straight, and under regular use it actually only barely reaches 51C for a few moments every now and then.

 

[Zap]

definitively, that undervolting in no way damages or reduces the lifespan of the CPU, yes?

It might.

Anyone remember the experiments at the now defunct DFI Street forums? They (someone or some few) postulated that if there was a big difference in voltages from RAM to CPU that the CPU's integrated memory controller will die (effectively killing the CPU). Several people tried it, basically by overvolting the RAM and undervolting the CPU. Results were that CPUs did in fact die.

Similar thing with the RAM voltage limit in the Core i7. 1.65v is the limit if the CPU is at stock voltage, but supposedly that can (somewhat) safely be increased beyond 1.65v if the Uncore voltage gets raised. This points to the problem being the difference between the two, not the actual voltages.

 

[jvroig]

That's an interesting post, Zap.

RAM voltage is what, 1.8V? If it is the difference between the RAM and CPU voltage instead of just the actual voltages, then raising the CPU to 1.8 won't be harmful? That is if it is taken literally that the wider the gap between the two, the more harmful it is between the CPU.

Or perhaps what is meant is that there is an "ideal gap" between the two, depending on the CPU? Say, for example, for old Athlon X2 chips the ideal voltage gap is .5V, and for other families of chips it may be higher or lower, and whenever you widen/lessen that gap significantly (undervolting or overvolting your CPU)) you risk damaging the CPU?

After reading so a dozen or so posts on other forums about undervolting, this is the first time I have heard of this. I cannot really comment more, as this is well beyond my area of expertise. I hope some others can chime in and say their piece.

 

[soccerballtux]

That's an interesting post, Zap.

RAM voltage is what, 1.8V? If it is the difference between the RAM and CPU voltage instead of just the actual voltages, then raising the CPU to 1.8 won't be harmful? That is if it is taken literally that the wider the gap between the two, the more harmful it is between the CPU.

Or perhaps what is meant is that there is an "ideal gap" between the two, depending on the CPU? Say, for example, for old Athlon X2 chips the ideal voltage gap is .5V, and for other families of chips it may be higher or lower, and whenever you widen/lessen that gap significantly (undervolting or overvolting your CPU)) you risk damaging the CPU?

After reading so a dozen or so posts on other forums about undervolting, this is the first time I have heard of this. I cannot really comment more, as this is well beyond my area of expertise. I hope some others can chime in and say their piece.

well it's designed to handle 1.8v though. All this stuff is up in the air if it's designed for the voltage range in mind. For example, Silicon TRIACs handle + and - 120VAC (which peaks at (2)^(0.5) times 120V).

So the damage would come from the gap being greater than the chip was designed for.

For what it's worth, the mobile chips are usually just best-of-the-best of the wafer; ie the chips that will run at the lowest power (different architectures aside). So I think you'd be ok with minor undervolting-- ie as low as the processor can go-- since desktop chips won't let you undervolt all that far.

My "slow" speed to save power is 1.07v and 11x200 multi.

 

[jvroig]

After lowering the clock speed by 800 MHz (13x multi, the equivalent of a Phenom II 910), I was able to undervolt even lower and still be stable, at 1.168V (previously, at full speed of 17x multi, would not be stable at 1.186V).

Temps improved by 1-2C at idle, and 4-5C at full load (Prime95). My CPU is actually pretty cool running after the initial undervolting, so it's not really the temps I'm after right now. I just found myself with too much power for the moment, since this desktop is temporarily mostly on all the time running Transmission BT client, and serving as a very overpowered typewriter (running OpenOffice Writer and Calc) until I finish my thesis this month (must not disappoint my employers, they are footing the bill for my master's degree). For now, not much other work crunching data and generating reports, and much less any gaming, so I decided to underclock and undervolt some more.

So the damage would come from the gap being greater than the chip was designed for.

I'm curious why that would be the case. Has anybody (or maybe even official word from Intel/AMD) why this is so? Why would the voltage discrepancy between the memory and CPU have any effect at all on the CPU's lifespan?

 

[richierich1212]

When you undervolt you also have to play around with the NB voltage and try to find stability.

 

[jvroig]

When you undervolt you also have to play around with the NB voltage and try to find stability.

I was not aware of this, thank you. Do you mean I also have to trying overvolting and undervolting the NB voltage to see what makes the CPU undervolting stable? Or is there a general rule of thumb that says how much more to over/undervolt the NB when undervolting the CPU?

 

[heyheybooboo]

...


For the first time in a build, noise became a problem as the Phenom II X4 stock cooler was quite unlike the ones in the old Athlon X2 5000 and even the 7750. It whirs like crazy at full speed (6750 RPM), and even at 5000RPM it starts to be very noticeable and annoying.

....

That seems a little wacky.

There may be a sensor or BIOS issue at work causing such high revs.




--

 

[jvroig]

There may be a sensor or BIOS issue at work causing such high revs.

As far as I can tell, 5000RPM happens under load and is normal (Bios smart fan enabled). 6750 (max RPM) only happened when I tested what the max is (through linux). 6750 is noisy like crazy, but even 5000 RPM is already annoying, especially at night when things are a lot quieter.

I've also read a recent thread here where another Phenom II X4 owner complained about the stock cooler being rather noisy, starting at around 4800 RPM. So I guess the stock cooler really is just noisy whenever temps start to reach ~42C maybe.