Pushing an i7 920xm to new levels by increasing the CPU core voltage

[svl7]

This post will quickly describe a way to increase (and decrease) the core voltage of the CPU, I have to advise anyone against doing this unless you're aware of all the risk that is involved.

First of all a quick summary of how the voltage regulation works in the recent Intel CPUs, if you want to see my results skip to the end of the post.

The method described below is mainly interesting for unlocked extreme processors, it will allow you to push the multiplier a bit more without crashing the systems.

If you happen to have an adjustable BCLK it may be helpful as well and should allow you to push it further than with the standard voltage (provided the RAM can keep up with the speed)

A quick explanation: The core voltage (Vcc) gets regulated by combining 7 different voltage identificators pins (VIDs), VID0, VID1, ... VID6.

Each VID has a different voltage value.

VID0 = -0.0125V

VID1 = -0.0250V

VID2 = -0.0500V

VID3 = -0.1000V

VID4 = -0.2000V

... you get the idea.

Simplified you can say the Vcc is at 1.5V when no VID is active, every active VID pin will subtract it's value from the Vcc and in this way create the needed Vcc.

Looking at the VID table should help to understand the concept. A 1 represents an active VID, a 0 an inactive VID pin.

The seven VID pins allow a wide variation of different voltages which allow smooth transitions and can get easily controlled with a simple binary code.

E.g. the string for 0.9250V would be 0101111

Now if you permanently deactivate a VID pin by connecting it to Vss (CPU ground), you increase the voltage of each possible VID combination which uses this pin actively (set to 1).

This means with this method you always change a whole range of different voltages.

For example if you connect VID3 to ground, you will increase each voltage which contains VID3 = "1" in the table by 0.1V.

The problem is that this will result in voltage jumps when the CPU adjust its voltage.

As an example lets assume VID3 is connected to Vss (this means VID3 is always = "0" in the table)

Now if you look at the VID table, 1.1000V is a VID combination of 0100000. If the CPU increases the voltage by 0.0125V it will set the VIDs to 0011111 which normally results in 1.1125V, but since the VID3 is always set to "0", the voltage will directly jump to 1.2125V.

Intel actually requires a smoother voltage transition, only 12.5mV steps are allowed according to the datasheet.

This voltage mod will result in a "jumping" Vcc each time the modded VID pin should get set to "1". So this mod won't give you a voltage increase in a percentage basis, but only increase it at certain VID combinations.

Therefore it's (in my opinion) only suited for increasing the max voltage, or (with some adjustments) also for decreasing the idle voltage.

Anyway, this will result in voltage changes that aren't conform with the Intel standards.

Also if you plan to do such a mod, I highly recommend to measure the voltage of your CPU before doing this.

Intel says that "Individual processor VID values may be set during manufacturing so that two devices at the same core frequency may have different default VID settings", this means that the max voltage of your processor may be different from my value.

After measuring the max voltage which my 920xm uses when pushed to it's stable limits (multiplier at 26x at all cores, baseclock at standard 133Mhz), I took a look at the VID table in the datasheet of the processor.

My max voltage was at about 1152mV. This means by setting the VID4 to "0" (by connecting it to Vss) it gives me a boost of 200mV when my CPU gets to its limits.

The datasheet provides all the necessary information about the pin locations. Below you see how it looks, encircled in red is what must get connected to permanently set VID4 to "0".

For doing this you simply connected the Vss and VID4 pin in the CPU socket with a very thin wire. Make sure the CPU pin has still enough room.

If it's done it'll look like this:

I measured the voltage to make sure it worked as intended. The left picture shows the voltage a multi of 26x on all cores before the mod, the right pic the voltage after the mod.

I didn't do a lot of testing yet, but the results are very impressive so far. I've been able to raise the multi to 30x on all cores an run a quick TS bench. And with a multi of 28x and a BCLK of 144 MHz I was able to run the 32M TS bench at more than 4GHz on 8 threads.

The previous limit was 26 on all cores.

Left: Check out the multi shown in Throttlestop

Right: 32M TS bench at 4GHz on 8 threads

I'll see what I can do with proper cooling... this is gonna be fun

_______________________________

A voltage increase will result in an extreme increase of the power draw, resulting in a lot of heat.

I have to advise everybody against doint this unless you're really aware of the risks involved in this.

This is nothing that you should do to your machine permanently, but it's great to see how far it can be pushed. It's an extreme kind of overclock which will stress the system very hard.

The Intel "absoulte maximum rating" for the voltage of the 920xm is 1.45V, so this is really close...

_______________________________

As already mentioned, this method can also be used to decrease the idle voltage. You'll need to find out your idle voltage and then check the VID table for the suitable mod.

You'll have to connect the VID pin(s) to Vcc instead of Vss if you want to decrease the voltage (set the VID permanently to "1")

There are other possibilities to increase the voltage, also to get a certain percental voltage increase which won't have the problem of jumping voltage values. I might investigating in this sometimes. The advantage of this mod is, that it's very simple. No soldering iron is needed and it can be quickly removed.

______

Let me know if something is unclear, this is quite complex to understand and I wanted only to quickly explain the idea behind this.

I'll be glad to help you in case you want to under- or overvolt your CPU with this method and have some questions.

PS: I'll try to put this in a frontpage article when I find time.

Edited July 12, 2011 by svl7
found some typos, there are probably more , haha

[StamatisX]

@svl7 Excellent post as usual

What I have noticed is that unless you have very good cooling (air-condition,dry ice,etc) the voltage increase produces too much heat like you mentioned on your post and you get the opposite results like lower scores on benchmark because the multipliers drop in order to keep the temperatures down, but this is the only way if you want to overclock beyond what BIOS can provide you and set a few records

[svl7]

Thank you! Yeah, this is only for hardcore benching, with additional cooling... and it stresses the hardware very hard. But it's great having the 920xm at 4GHz on all cores

[StamatisX]

when you run wprime 1.55 can you maintain those multies?

[svl7]

Hmm... apparently no. Shutdown and PSU light is out. Have to un- and replug the PSU from the wall to get it working again... Will have to investigate a bit more in this. As I said, haven't done any intense testing so far, only some quick TS with crazy values.

Guess it's a power issue, never had such shutdowns before. Should try it with a bigger PSU, don't have one though. But I don't think it will help.

I'll let you know when I have more results. According to a friend of mine wprime 1.55 uses the CPU in a special, outdated way... It's weird that TS Bench works and wprime apparently not. I'll do more testing.

[StamatisX]

You can try IntelBurn Test or wprime95 for more load to the CPU (but it is even more unrealistic than wprime) and yeah the TS bench is lighter than the wprime 1.55

[Stevenxowens792]

That is seriously pushing hard. Great numbers!

[unclewebb]

TS Bench is a mild load in comparison to wPrime or Prime95. I know Prime95 uses some highly optimized assembler code that can take advantage of what these CPUs are capable of. If you have a Kill-a-Watt meter handy, that's an interesting way to compare different benchmark programs and separate the men from the boys. TS Bench is in the boys category.

Congrats svl7 on pushing the envelope. A 30X multi might not be practical for 24/7 use but it looks like a lot of fun. The 45nm Core i7 series is very robust so I don't think you'll hurt your CPU playing with it. Many users on XtremeSystems have pushed very similar desktop CPUs to well over 4000 MHz with lots of voltage and have rarely had a problem. There are less reports of these degrading compared to the newer 32nm CPUs.

[svl7]

... wow, I just some quick tests with the kill-o-watt plugged in, during a wprime 1.55 at 26x it peaked at 198W... I'm pushing the PSU harder than the CPU.

Kill-o-watt isn't that accurate of course... it's between wall and PSU so there's also the power draw of the PSU itself in the value (plus measuring error), but at least it delivers some values to compare.

Will probably try a 100mV mod sometime later instead of the 200mV... this should lower the power draw by about 20-30W.

[iloveb00bs]

holy crap bro, that friggin amazing. really wish i still had my M15x again just to blow it up hahahah +rep when allowed.

[StamatisX]

Killawatt reported 197W for TS bench

and 193W for wprime 1.55

never the less wprime 1.55 pushed the CPU harder (the multipliers dropped lower than TS Bench) even though the power consumption was less (4W less to be accurate)

I am with 75mV extra, 5% BCLK overclock and 1.6V for the memory

[svl7]

Killawatt reported 197W for TS bench

never the less wprime 1.55 pushed the CPU harder (the multipliers dropped lower than TS Bench) even though the power consumption was less (4W less to be accurate)

I am with 75mV extra, 5% BCLK overclock and 1.6V for the memory

The only time I raised the RAM voltage I ended in an infinite boot loop, without seeing the POST screen anymore...

What's your multiplier limit with the 75mV VCC increase? I can run stable at 26x without voltage adjustments (tested with wprime 1.55 1024M, vantage, 3dM11).

[unclewebb]

StamatisX: I think I had some similar results when testing on a Core i5 desktop but couldn't figure out why something that consumes less watts can result in more turbo throttling. When designing the TS Bench, I had one eye on the Kill-a-Watt meter and I kept adjusting the algorithm by adding some ugly looking math functions until I saw a big number on the Kill-a-Watt.

Very scientific.

The Kill-a-Watt is not 100% accurate but the one I have seems to be fairly consistent from day to day. Gives you a rough idea how much torture you are putting your PSU through.

Did you try any quick Prime95 or LinX testing for some comparison numbers?

Edited July 12, 2011 by unclewebb

[StamatisX]

StamatisX: I think I had some similar results when testing on a Core i5 desktop but couldn't figure out why something that consumes less watts can result in more turbo throttling. When designing the TS Bench, I had one eye on the Kill-a-Watt meter and I kept adjusting the algorithm by adding some ugly looking math functions until I saw a big number on the Kill-a-Watt.

Very scientific.

The Kill-a-Watt is not 100% accurate but the one I have seems to be fairly consistent from day to day. Gives you a rough idea how much torture you are putting your PSU through.

Did you try any quick Prime95 or LinX testing for some comparison numbers?

My guess is that it also depends on the programming language used in each program and how multithreading was implemented, if wprime 1.55 uses assembly for instance, it will utilize the CPU more efficiently than a program being written in C++ or C#. So I suppose the CPU does not idle as much and it requires slightly less energy but it heats up faster that's why the multis drop more as well. I can't think of a different explanation at the time being.

I will try prime95 right now and I will post the results. I know for a fact that at those clocks it will give me a nice BSOD

The only time I raised the RAM voltage I ended in an infinite boot loop, without seeing the POST screen anymore...

What's your multiplier limit with the 75mV VCC increase? I can run stable at 26x without voltage adjustments (tested with wprime 1.55 1024M, vantage, 3dM11).

mine is 26 x 140.3 (3,6489 GHz) I use it for everyday since no program (including those you've mentioned) pushes the CPU that hard for that long in order to crash. It's all about ambient temperatures though, if your room is airconditioned then it's fine, if not...

[StamatisX]

With prime95 at those clocks I don't get a BSOD but I got lots of erros, I am not stable and killawatt reported 215W...

[mw86]

So is prime95 the hardest bench to pass and the greatest load on the cpu from all benching software? You know something interesting with prime95... the long term runs push things to the limits... if in prime95 you hit bench or something it starts loading the cpu on single threads adding more and more threads of usage and my system was stable on small ffft and all those tests but the bench ran about 3-4 times in a row immediately after the other really told me if I was stable especially if I had set different multiplier limits per core usage... this doesnt apply to your usual 4 core constant max turbo tweak.

By the way SVL7 that is one hell of a mod. I miss my 920XM already. With no load I've had the 920xm around nearly 4.25ghz plus others Im sure have surpassed even that and I think using this type of mod Svl7 you could get there on a bench.

Edited July 16, 2011 by mw86

[Jimbo]

As far as I'm concerned prime is one of the big two, linpack (intel burn test) being the other. They're both totally brutal on CPUs.

[mw86]

I love the Intel Burn test its my favorite at max ram usage... Prime is probably the top cpu burner out there.

[svl7]

I got my 240W PSU today, and some quick tests indicate that it indeed helps! I only did some very quick testing so far since I don't have sufficient cooling possibilites, but I managed some 32M wprime tests, not a single crash I pushed up to 28x on all cores and 143 Baseclock... hehehe. PSU peaked at 210W according to my kill-o-watt, the regular M15x PSU tripped instantly at about 200W.

6.414 seconds so far... more than 4GHz on eacht thread, though I guess some of the programs and processes running in the background slowed it down a bit. I'll do some real benching soon and then I'll see how much further I can push this.

[StamatisX]

Nice! do you have any external cooling like an AC unit or is it room temperature?

[svl7]

This was with room temps, no cooling pad, I didn't even remove the backplate... it doesn't matter for such a short time. It may peak at 95°C, but that's ok.

[StamatisX]

Give it a try with the 1024M, probably you will have to lower the multis to x25 or x26.

[svl7]

I will as soon as I have cooling, it won't work without additional cooling as long as I have the voltage mod applied... temps go crazy with a 200mV higher core voltage. It would throttle immediately, or, with Throttlestop activated, hit 100°C and shut down.

I can do 26x without voltage mod, this results in about 222 seconds.

[StamatisX]

I have mine with the backplate removed and sitting on the cryo lx at 75mV and 140x26 with 98/96 TDP/TDC with temps maxing out at 97 C and it does the 1024M in 240sec. I need to try with an air-condition blowing some cold air from underneath...

*EDIT*

Same with vantage

[svl7]

That's pretty good, 240 secs means that there's only very slight throttling at the end of the test. My 222 seconds with 26x were done with very low ambient temps, no throttling. With room temps and 26x I got about 260 seconds if I remember correctly... but as I said, the voltage increase causes the temps to go crazy, no long runs without auxiliary cooling.