Core i7 Vdimm/VTT Graveyard :D

[cadupradoo]

Humm it sounds weird that you needed to set the IOH higher just like that...

Which BCLK are you running at?

[saaya]

1.4v is high indeed, so far 1.3v seemed to be enough for everybody i heard of...
reboots huh? hmmmm what are the cpu and chipset temperatures? what psu are you using?
do you use eist? did you try to bump up vcore to see if that helps?

[Amurtigress]

Hi saaya,

my current emps are roughly 37C on the chipset/board and core temps are about 71C in Realtemp or 61C in Speedfan. Stupid unclear Tjmax specs... Running the memory at 1200 doesn't change the temps much, if at all.

Not using EIST. All CPU features except for the prefetchers are disabled.

My PSU is a Coba Nitrox 750W. It is a OEM PSU made by CWT, it has the same 'innards' as the Thermaltake toughpower series with their fake four 12V rails. I was surprised to see a double-sided green PCB inside instead of the normal single-layered ones.

Matter of factly the 12V rails are connected together internally.The MOSFETs and diodes for the 12V circuit are specified for 80 amps. Makes sense if the total current for 720W on the 12V rail alone could be up to 60A. It is one year old and I checked the elcaps when I put the i7 together. All looked fine. Voltages are stable with my multimeter. I can't see ripple tho without an oscilloscope.

As for my core voltage, 1.25V are enough for 3.8 GHz. I am currently at 1.30, but I did not intend to go further than that.

I read a recommendation from TheRAMGuy at the Corsair forums to someone who had an issue that sounded much like mine, but with 6 modules. RAMGuy recommended to set the QPI voltage to 1.5, but that did not change anything for me.

I've tried IOH voltages of 1.20, 1.26 and 1.36-1.4. Anything above 1.26 did not show a change, 1.20 was unstable. So 1.26 seems to be about the sweet spot.

Hmmm....

[saaya]

qpi voltage is vtt, it actually has nothning to do with qpi, its powering the uncore of the cpu, the l3 cache and imc... so yes, if you use a lot of mem at high clocks or tight timings you need to push the uncore voltage up so the imc can run stable... ioh is the northrbidge, the x58 chipset, increasing ioh 1.1v results in higher and higher stable bclocks, its not related to memory directly.

so what exactly is your problem? getting mem clocks up?

[cadupradoo]

Like saaya said, qpi voltage is "only" about everything in the processor, but the cores... So it's pretty important for memory overclocking..

I wish evga x58 bios offered a 9x multiplier for the memory... I can't use 2000MHz with the 200BCLK, but I wish I could use 1800 instead of 1600...

[saaya]

as far as i know there are no uneven memory multipliers
even on 1156 there wont be... thats where unlocked cpu multis really come in handy, you have a much wider range of cpu and memory speeds you can run... but getting a 965 just for that doesnt make sense, of course...
i asked intel for half multipliers or at least uneven multipliers a while ago but they said they dont think its necessary...
more multipliers always mean a slight loss in efficiency and can cause annoying bugs, need more time to test as well to make sure they all work...
in the end intel is right, especially seeing how extra mem clocking barely makes a difference... and with ddr3 we are quite flexible with timings and speeds, so just tighten the timings id say...

1600 cas6 is way tight and as fast or faster than 1800 cas7...

[Amurtigress]

qpi voltage is vtt, it actually has nothning to do with qpi, its powering the uncore of the cpu, the l3 cache and imc... so yes, if you use a lot of mem at high clocks or tight timings you need to push the uncore voltage up so the imc can run stable... ioh is the northrbidge, the x58 chipset, increasing ioh 1.1v results in higher and higher stable bclocks, its not related to memory directly.

so what exactly is your problem? getting mem clocks up?

I can only repeat myself:

I am running my i7 920 at 19x200 and either DDR3-RAM clock 1200 or 1600. My memory is 4 pcs./2 sets Corsair TW3X2G1600C9DHX modules that I bought in fall 2008 with the perspective of using them on my future Nehalem system. Memory sockets A1, A2, B2, C2 are populated according to Intel's specs.

My modules are working fine at 1200/7-7-7-20, 1.5V.

However, 1600 is a bit of a different story.

1600/9-9-9-24/1.8Vdimm and 1.35-1.5V QPI/uncore leads to random sudden reboots of my system without a bluescreen.

I am not getting any errors in LinX, no errors or failing worker threads in Prime 95 x64, and no errors in Memtest 86 plus V2.11 either. The temperature of the cores and chipsets are the same no matter if I am running 1200 or 1600 memclock.

I can certainly post a filled out template for my P6T Deluxe V2 board if neccesary.

PS: A quick thought, isn't it better to rather use the module in socket A1 in C1? I know A1 is the correct configuration, however, poulating C1 instead would make sure to have the on die termination for the black sockets at the end of the board's traces.

[saaya]

sorry, i dont get what you mean with swapping the one stick to another slot...

your mem cant run 1600 999? what kind of rating is that anyways? 1600 999 is teribly slow...
so your running 4 2gb modules, right? how about running 3x2gb, maybe that works? might be a bios problem since the bios remapps the memory address space...
might be worth giving a try, and i doubt you need more than 6gb... but 1600 999? damn thats terrible... isnt 1200 777 faster than 1600 999?

if you shoot me a pm with all your settings ill see if i can help you get it stable at higher speeds

[Amurtigress]

sorry, i dont get what you mean with swapping the one stick to another slot...

your mem cant run 1600 999? what kind of rating is that anyways? 1600 999 is teribly slow...
so your running 4 2gb modules, right? how about running 3x2gb, maybe that works? might be a bios problem since the bios remapps the memory address space...
might be worth giving a try, and i doubt you need more than 6gb... but 1600 999? damn thats terrible... isnt 1200 777 faster than 1600 999?

if you shoot me a pm with all your settings ill see if i can help you get it stable at higher speeds

I'll PM you shortly as soon as I've fetched all BIOS settings. My modules are 1 GB modules. The TW and 2G in Corsair's article codes refers to a twin kit of a total of 2 GB. So I wouldn't like to run on only 3 GB.

According to Everest 5.01, I am getting a read throughput of roughly 14 GB opn 7-7-7-20/1200 but up to 17 or 18 at 9-9-9-24/1600. The question is indeed if that would reflect in a higher system performance....

1600 9-9-9-24 is the XMP profile's settings stored in the modules, but they run fine at 8-8-8-20/1600 too, except for certainly still having those reboot issues.

What I mean by moving one memory module is that the manual states that for a four module configuration one should use slots A1-A2-B2-C2. '2' are the orange slots, and A1 would be the black one closest to the CPU socket.

When it comes to bus termination, and the RAM modules' on-die termination, wouldn't it make more sense to have the fourth module in the last of the black slots? To terminate the signals at the end of the copper traces.

[eva2000]

qpi voltage is vtt, it actually has nothning to do with qpi, its powering the uncore of the cpu, the l3 cache and imc... so yes, if you use a lot of mem at high clocks or tight timings you need to push the uncore voltage up so the imc can run stable... ioh is the northrbidge, the x58 chipset, increasing ioh 1.1v results in higher and higher stable bclocks, its not related to memory directly.

so what exactly is your problem? getting mem clocks up?

actually pushing ioh volts does help with mem clocks for linx/prime and super pi 32m stability http://i4memory.com/f80/i7-920-d0-df...ability-16617/ for some

[before]

2.10v vdimm / 1.82v vtt = ok for 2 days of non stop 3D Mark. (air)

[saaya]

actually pushing ioh volts does help with mem clocks for linx/prime and super pi 32m stability http://i4memory.com/f80/i7-920-d0-df...ability-16617/ for some

yeah, but only at high bclocks right?
when the prob is that qpi/bclocks arent fully stable, at least in my experience... and its a bit weird, some people seem to need high ioh, some say it doesnt help and some say it makes things worse if ioh is too high... i suspect it has to do with what bclock your running and possibly on the cpu and chipset.

btw, that zoom feature on large images is very nice...

2.10v vdimm / 1.82v vtt = ok for 2 days of non stop 3D Mark. (air)

thats high!
did you notice different ocing after this?
do you need more volts for the same clocks or can you reach only slightly lower clocks when maxing everything out?

I'll PM you shortly as soon as I've fetched all BIOS settings. My modules are 1 GB modules. The TW and 2G in Corsair's article codes refers to a twin kit of a total of 2 GB. So I wouldn't like to run on only 3 GB.

According to Everest 5.01, I am getting a read throughput of roughly 14 GB opn 7-7-7-20/1200 but up to 17 or 18 at 9-9-9-24/1600. The question is indeed if that would reflect in a higher system performance....

more bw helps in some apps, lower latency helps in others... the difference is really small in my experience, but games tend to prefer latency since bw on i7 is pretty massive to begin with.

1600 9-9-9-24 is the XMP profile's settings stored in the modules, but they run fine at 8-8-8-20/1600 too, except for certainly still having those reboot issues.

hmmm could be that the memory pwm gets too hot or cant cope with the load... but that doesnt make much sense since your on a p6t and they have 2 or 3 memory pwm phases right? do they get hot? ask asus tech support about it... maybe your on an early pcb rev and theres some prob... who knows...

i remember that the renessance had similar probs early in development, it was some prob with the memory pwm... but it only happened with certain psus for some reason... so kinda a psu mem pwm compatibility... try another psu, maybe its the same thing.

What I mean by moving one memory module is that the manual states that for a four module configuration one should use slots A1-A2-B2-C2. '2' are the orange slots, and A1 would be the black one closest to the CPU socket.

When it comes to bus termination, and the RAM modules' on-die termination, wouldn't it make more sense to have the fourth module in the last of the black slots? To terminate the signals at the end of the copper traces.

each channel is terminated independantly, so i dont think it matters... channel a has the shortest traces so thats why intel recommends using 2 dimms there if you use 4 afaik.

[Amurtigress]

Saya (and all others),

thanks for the reply.

I might have shed some more light on my issue here. I came up with the idea to dig up the old and IMHO untrusty OCCT 1.1, but it seems like it shows some detail about my system. It might have to do with differential amplitude settings:

I tried the following (so far diiferential amp has always been AUTO):

memlock/ucore clock/D.A. -- stability

1600/3200/auto -- OCCT 1.1 stops after ~3 minutes
1600/3200/700 mV or 800 mV -- stops after 19-21 minutes
1200/3200/700-800 mV -- stops after ~3 minutes
1200/2400/700-800 mV -- stops after ~3 minutes
1200/2400/auto -- stops after 20 minutes.

(Remark: In all cases when OCCT failed after ~3 minutes it would also fail to update it's clock every second, but only every few seconds. It would be absolutely smooth in all cases when it bailed out after 20 minutes. irregular clock performance looks to me like it was a symptom of internal problems of my hardware)

The table above leads to an assumption: At 1200 the differential amplitude must be set to AUTO, and at 1600 most likely to anything around 700-800 mV. In the past OCCT 1.1 has been extremely flakey on my previous Q9550/DDR3 system, so I am tempted to dismiss it's failures after 20 minutes as issues with newer hardware.

Now testing my theory with OCCT 3.0.1 in OCCT mode:

1200/2400/auto -- 1h PASSED!
1600/3200/700-800mV -- in Progress

Saaya:

My board is actually the P6T Deluxe V2. The PCB is identical to the P6T Deluxe, with only a sticker on the original board model print on the board itself. The revision is 1.02G as on all shipped Deluxes...ASUS simply just left out the SAS controller chip and it's SATA connectors.

The P6T series has 16+2 phases, the two extra ones appear to be for the uncore voltage. I just added a fan to my northbridge since the heatsink and heatpipe were getting very hot when running the uncore at 3200 MHz (neccessary for DDR3-1600). However it was significantly colder at 2400/1200.

Cross fingers, I might get to the skinny about the issue!

[before]

thats high!
did you notice different ocing after this?
do you need more volts for the same clocks or can you reach only slightly lower clocks when maxing everything out?

Oh no, just tried to shot this sucker down

[Gautam]

2.10v vdimm / 1.82v vtt = ok for 2 days of non stop 3D Mark. (air)

[dinos22]

2.10v vdimm / 1.82v vtt = ok for 2 days of non stop 3D Mark. (air)

lol and i thought i was pushing a bit with 1.7-1.75 VTT

[eva2000]

dam 1.8+ CPU VTT Never tried above 1.65v myself for CPU VTT heh

yeah, but only at high bclocks right?
when the prob is that qpi/bclocks arent fully stable, at least in my experience... and its a bit weird, some people seem to need high ioh, some say it doesnt help and some say it makes things worse if ioh is too high... i suspect it has to do with what bclock your running and possibly on the cpu and chipset.

btw, that zoom feature on large images is very nice...

Not solely for high bclk but mem clock too but helps with anything that stresses all cores versus single threaded apps. But yeah each cpu will respond differently.

8x 32M Super Pi instance via HyperPi @DDR3-2080Mhz 7-8-7-20 1T at 1.725v bios set vdimm = 1.730v real DMM vdimm.

[road-runner]

I have been running 1.88 vtt for awhile now 24/7 crunching...

[dinos22]

I have been running 1.88 vtt for awhile now 24/7 crunching...

can i ask why
that seems a bit ridiculous if you dont really have to have it so high

[road-runner]

can i ask why
that seems a bit ridiculous if you dont really have to have it so high

Sorry not vtt I meant Pll...

[dinos22]

oh i see

[Chicken Patty]

oh i see

i've never had CPU PLL over 1.8v. I can use that all the way up to 4.3 GHz HT on. Wonder if the hard time I have to get over 4.3 GHz is due to not bumping up the PLL higher??

[Amurtigress]

Hi people, Saaya,

I am pretty certain now that my reboot issue is solved now. I had the idea the other day to look into settings for ASUS boards and certain memory modules they sell, tho not neccessarily the same as mine.

Some Corsair DDR3 modules require 1.9V to run stable at 1600 in two pairs in older systems, so I tried this. For some 1866 modules Corsair recommends to set uncore/QPI to 1.6V, which I did too.

The third factor is that to have 1.6V uncore working it would require a fixed Differential Amplitude setting of 700 or 800 mV.

My system passed seven hours of Prime 95 x64 the other night.

1.9V vdimm and 1.6V uncore seem a bit high and pretty much heat up the processor, but it doesn't reboot on me anymore.

I also came across this older german report on ASUS presumably having found ways to design the P6T boards in a way to allow for high VDIMMs without affecting the IMC. I have no idea if this holds true, also since this report was released on Oct. 8th. However they also refer to similar statements from Fudzilla at that time:

http://www.hardware-aktuell.com/news...m_limitierung/

Any comments on this?

Edit: I just found the Fudzilla article. However Fudzilla says nothing about this workaround NOT having an effect on the IMC.

http://www.fudzilla.com/index.php?op...9824&Itemid=37

Earlier article:
http://www.fudzilla.com/index.php?op...9489&Itemid=37

Let me try to sum it up what might have happened to let this 'myth' grow:

I believe that Intel's reference design for Socket 1366 mainboard was having one voltage rail that provides power to the uncore part of the i7 CPUs, and also to the DIMM sockets.

This would mean, if Intel said that more than 1.65V on the uncore could kill it, it would limit the VDIMM voltage along with it.

However, mainboards that have a separate VDIMM voltage setting in their BIOS, and a separate VDIMM switching regulation, would not have this limitation on the memory modules.

Or, to put it differently:

The UNCORE is limited to 1.65V, NOT the memory voltage. UNCORE above 1.65 can kill your processor, and most likely NOT VDIMM.

Then came this stupid sticker on the ASUS boards that began to feed the legend of high vdimms killing CPUS. And we know now from Saaya that only a small number of excessively high voltages on modules during the board design were actually killing a smaller number of A and B stepping engineering samples at Foxconn.

Per chance noone ever said VDIMM above 1.65V would kill the IMC (save for ASUS' sticker), unless they meant VDIMM and QPI/uncore voltage were THE SAME.

OK, this is just a theory based upon what little I know about reference designs, but it seems to make sense.

Feel free to flame me now.

[Bei Fei]

Has anyone killed a CPU by running 1.45 QPI 24/7?

[Hondacity]

Has anyone killed a CPU by running 1.45 QPI 24/7?

thats normal, i'm using 1.7v on the qpi and it isn't dead yet