QPI Voltage for DDR3-1600 CL7

[madrooster]

I've got a 1600MHz CL7 kit here which works with the following on a C0/C1 i7 920 at stock CPU freq (for now) on a Gigabyte EX58-UD5:

10x RAM multi to get 1600
Vdimm 1.65V
Vcore auto
V-IOH auto
V-QPI 1.34V

All other voltages on auto. The kit works fine for 25 runs of LinX on max problem size, however when paired with a D0 stepping i7 920 on the same motherboard as the C0/C1 CPU, it refuses to run with the above voltages (errors within 5 mins).

The kit, with the D0 CPU, will run at 1333MHz CL7 with Vdimm 1.5V and V-QPI 1.175V no issues.

Am I missing anything? Do I need to adjust the QPI PLL too?

 

[mastrdrver]

Remember that since the i7 has the memory controller on the chip, no 2 chips will work at the same settings given the same hardware. You need to start over and relearn your D0.

Since it errors in 5 minutes, usually, that indicates that you need more vcore. Have you tried giving it more?

 

[Binge]

I've got a 1600MHz CL7 kit here which works with the following on a C0/C1 i7 920 at stock CPU freq (for now) on a Gigabyte EX58-UD5:

10x RAM multi to get 1600
Vdimm 1.65V
Vcore auto
V-IOH auto
V-QPI 1.34V

All other voltages on auto. The kit works fine for 25 runs of LinX on max problem size, however when paired with a D0 stepping i7 920 on the same motherboard as the C0/C1 CPU, it refuses to run with the above voltages (errors within 5 mins).

The kit, with the D0 CPU, will run at 1333MHz CL7 with Vdimm 1.5V and V-QPI 1.175V no issues.

Am I missing anything? Do I need to adjust the QPI PLL too?

What is the core speed you're running with this overclock? It probably isn't a memory problem if you're even booting.

 

[madrooster]

On the C0 stepping chip, I could leave the vcore on auto with the usual power saving features on (basically defaults, stock core freqs), but with the RAM running the sticks' rated 1600MHz CL7. There is no core overclock in this case, only the uncore (as a result of the 1600MHz RAM).

On the D0, same settings. Stock core frequency (2.66GHz) but with the uncore on 3200MHz. I haven't got time to fiddle with the core side of things yet, so for now I just want to run the 1600 kit at its rated speed. However the D0 is the only difference between the 2 setups.

I have now tested running the kit as 1333MHz CL7 - no issues. 1600MHz CL9 - bombs on the 2nd run of LinX. 1600MHz CL8 and CL7 - bombs before it finishes the 1st run of LinX.

 

[mastrdrver]

On the D0, set the vcore equal to that of what your vvt is and see if that helps, with LLC disabled so that it follows intel spec.

 

[madrooster]

I'll try 1.2V vcore first, as that's what the C0 maxes out on, at full load with EIST, C1 etc. turned on.

 

[mastrdrver]

Stop thinking your D0 will be like your C0.

 

[Binge]

On the C0 stepping chip, I could leave the vcore on auto with the usual power saving features on (basically defaults, stock core freqs), but with the RAM running the sticks' rated 1600MHz CL7. There is no core overclock in this case, only the uncore (as a result of the 1600MHz RAM).

On the D0, same settings. Stock core frequency (2.66GHz) but with the uncore on 3200MHz. I haven't got time to fiddle with the core side of things yet, so for now I just want to run the 1600 kit at its rated speed. However the D0 is the only difference between the 2 setups.

I have now tested running the kit as 1333MHz CL7 - no issues. 1600MHz CL9 - bombs on the 2nd run of LinX. 1600MHz CL8 and CL7 - bombs before it finishes the 1st run of LinX.

Ok, so what is the furthest you've gone with your vtt/qpi voltage to make it stable?

 

[madrooster]

@ mastrdrver

I'm well aware of that, though generally D0's require less voltage compared to a C0 - not more!

@ Binge

I've only gone as far as "safe" voltages go, ie. 1.35V QPI/VTT, 1.65V VDIMM. Best I can do with those is 1333MHz CL7 on the D0.

On the C0, no issues running 1600MHz CL7 with those voltages. Vcore and all other voltages were not touched in either scenario.

 

[Binge]

@ mastrdrver

I'm well aware of that, though generally D0's require less voltage compared to a C0 - not more!

@ Binge

I've only gone as far as "safe" voltages go, ie. 1.35V QPI/VTT, 1.65V VDIMM. Best I can do with those is 1333MHz CL7 on the D0.

On the C0, no issues running 1600MHz CL7 with those voltages. Vcore and all other voltages were not touched in either scenario.

Your response to mastrdrver is not true. The maximum overclock for 920s went up with the D0, and likewise minimum voltages to attain 4.0GHz overclocks went down. All the while it still doesn't matter if you get one that sucks. My first C0 performs better than most D0 today. I'm sorry your hardware isn't behaving, but it's seemed like you have an equal understanding of this hardware than anyone here. By no means is that us giving up, but I also get the feeling you aren't waiting around for us to give you an answer. Every time you make a step forward or a step in a direction it would be to our benefit if you logged your attempts here.

A thought comes to mind. Was your motherboard's bios ever updated?

 

[madrooster]

No, the BIOS remained the same the entire time. Everything else has stayed the same. Only the CPU changed.

 

[Binge]

No, the BIOS remained the same the entire time. Everything else has stayed the same. Only the CPU changed.

There was new microcode introduced when the D0 became available. The lack of an update may be causing problems with c-states if you use them.

 

[madrooster]

The C states are enabled yes - though under LinX it should be in C0 state the entire time.

I'm running the F8 bios, pretty sure the D0 microcode was introduced ~F5.

I have now tried 1.2V Vcore, VQPI 1.34V, VDIMM 1.65V, CL7, 1600MHz. No go, it bombs after the 2nd run of LinX (it was bombing prior to finishing the 1st run previously with the Vcore untouched).

Now trying 1.3V Vcore.

 

[SummerDays]

I don't know whether this is worth mentioning or not, but if you're using an i7 you can set the ram frequency by just choosing the appropriate speed in the bios, and leave the rest of the settings on auto. For example, 1333 Mhz or whatever it is, and then set the QPI to 6.4T if you want. Then after that, run the i7 stock base clock speed.

It will produce a bit of a temperature increase on your cpu.

 

[Binge]

I hope it works out for you. I've had a D0 that wouldn't even post into windows at stock, so don't be surprised.

 

[madrooster]

@ SummerDays

No this is a VERY bad idea on (at least) the Gigabyte X58 boards. When you just set the multi to 12 and leave everything on auto, the board uses 1.8V VDIMM which is well beyond Intel's recommended spec. Similarly setting it to 10 (for 1333MHz), it tries to use 1.6V VDIMM for modules which are rated 1.5V. You MUST set voltages manually for anything > 1066MHz on the Gigabyte X58 boards.

@ Binge

Thanks!

Currently I have them stable on 1333MHz 6-6-6-20 1T on 1.6V VDIMM (stock QPI/VTT etc.) with 50 runs of LinX. I'm pushing for lower voltage now - trying 1.54V VDIMM.

Now if only SPDTool supported editing DDR3 SPDs.....

 

[Binge]

Have you tried to see if you can run a higher bclk with lower memory multipler @ 1600MHz?

 

[SummerDays]

Thank you for telling me that.

On the Rampage II Extreme, you just set the memory speed you want, and QPI up to 6.4 and it does the rest.

Intel didn't want to, but the memory manufacturers were upset that their memory would be forced to run at 1066.

 

[madrooster]

Well, I recently noticed that even the C0 cannot stay 100.0000000% stable running 1600MHz 7-7-7-24 2T. It errors around the 60th run of LinX.

On the C0, 1600MHz 9-9-9-28 2T seems to pass 100 runs of LinX.

On the D0, 1600MHz refuses to pass 5 runs even.

However *both* the C0 and D0 pass 100 LinX runs with the same settings for 1333MHz 6-6-6-20 1T. I think I'll settle at that.

For 1333MHz 6-6-6-20 1T:

QPI/VTT: Stock (1.15V)
VDIMM: 1.6V
Vcore: Auto
IOH: Auto
RAM Multiplier: 10x

Thanks everyone for their suggestions along the way.

 

[mastrdrver]

You don't need to run a lot of loops for LinX to find stability. The flop numbers will increase up to a certain point with voltage increases. This is where I stop and I've had zero problems with this method. Though, you need to make sure your vcore is good as not having enough will create problems. Not to mention that you won't make it more than 10 mins with too low vcore.

 

[madrooster]

Just an update on further fiddling I eventually got around to...

I found the culprit to LinX erroring within the first run whilst running as 1600 CL7 on the D0 chip. Turns out the tRFC timing that is set in the XMP profile is not appropriate at all for the 128MB chips the dimms use.

Even if you tell the board, 12x multi and set the right voltages without the XMP profile it would use a relatively tight tRFC timing which still wasn't appropriate.

Changing the tRFC to 88T immensely increases stability of these sticks running at 1600 CL7. I could pass anywhere up to 80 runs of LinX in one go.

So I am at a stage where the sticks run CL7 at a minimum of 20 runs of LinX.. a maximum of 80 odd runs in a single run, with 1.335V QPI and 1.64V Vdimm.

Suggestions on making it more stable? Would adjusting the QPI PLL do anything?