Intel 4.5GHz & 5GHz LinX Stable Club

[PiLsY]

Regarding GFlops: my main problem is that GFlops are going up and down with no change to voltage whatsoever... it's really annoying. Today I did two runs (at different times, with several restarts in between) with the exact same voltages of 1.4125 vcore/1.35 QPI, but I mysteriously lost 4GFlops somewhere along the line. Every single setting in BIOS was unchanged.

I went through the same pain with all of mine - i3, i5 and i7. It took a long time to crack it but the last 50mhz is always the hardest. Try messing with LLC and raising/lowering bios vcore to match load voltages. I always run LLC completely at intel stock for the last bit of stability - however asus boards behave somewhat differently so just experiment a bit. P8P67-Pro has made me break that rule .

Its probably worth taking a step back and finding the nearest stable setting and work from there. Diff bclk/multi maybe?

[zoson]

Your turn to step up, don.

[donmarkoni]

I'm coming for your crown, don.
So far i'm 8 loops stable 100GFlops at 4536MHz core, 4067MHz uclk, and 2190MHz mem.

I was about to say that it would be my honor for you to take it, when I saw your next post.

Your turn to step up, don.

http://i144.photobucket.com/albums/r...0-100GFlop.jpg

Congratulation!

[PiLsY]

Finally managed to haul my sandy into the club. Its literally the worst 2500k ever (4.9ghz 1.5v) coupled with a terrible p8p67-pro plus im running 4 dimms. Im really paying the price for my luck getting a good previous gen i5 and i7.

Anyway, this took literally months to get stable .

i5 2500k (batch to follow - got it written down but my notepads in work)
50 x 100.4 (5023.7mhz)
PLL OV - Enabled
LLC - High
Vcore - +0.085 (~1.536v load)
vdimm - 1.55v
vccsa - 0.950v
vccio - 1.1v
vcpu pll - 1.55v
vpch - 1.1v

Screenshot.

[Mr. K6]

Got some holiday cash and picked up a i5-2500K + Z68XP-UD4. I was surprised at how little there is to tweak . Currently running 5.0GHz as my daily overclock, here's the screenshot:

Settings:
100.2 x 50
Normal Vcore + 0.100V (~1.45 in BIOS, 1.404V under load in CPU-Z)
VTT 1.050V
CPU PLL 1.7V
CPU Overvoltage enabled
Turbo Disabled

Specs in sig

[LennyRhys]

Here's me struggling to get 85 GFlops at 4.5GHz, and you're nailing 100 GFlops at 4.53GHz - LOL!! Nice work

Thank you guys for your help and I'll see if I can find some time tomorrow to tweak vQPI, set my memory sub timings manually, and get the most GFlops that I can with 1800 7-8-7-24 and 3200 uncore. If I don't have any success I'll take Pilsy's advice and go for a different BCLK/multi.

[zoson]

Here's me struggling to get 85 GFlops at 4.5GHz, and you're nailing 100 GFlops at 4.53GHz - LOL!! Nice work

Thank you guys for your help and I'll see if I can find some time tomorrow to tweak vQPI, set my memory sub timings manually, and get the most GFlops that I can with 1800 7-8-7-24 and 3200 uncore. If I don't have any success I'll take Pilsy's advice and go for a different BCLK/multi.

Keep in mind, the above run I did was with HT off. I don't think I've ever gotten more than 88GFlops at 4.5GHz with HT on. In fact, I know the highest ever GFlops I've gotten with HT on was 91.5GFlops at 4657MHz core and 4000MHz uclk.

HT off and HT on are very different ball games.

Also, I went nuts and was running my uclk at 4067MHz, which requires me to set vQPI to 1.49375v in bios. :O Luckily, my chip seems to be no worse for wear.

Does your chip not do 3600 uncore with 1.35v QPI? Optimally you want your uclk to be 2x your memory speed. So 1800MHz mem and 3600MHz uclk goes like bread and butter. Bringing your uclk up to 2x mem will also probably net you ~2 GFlops

Four out of the five gulftowns I've owned did 3600 uncore with 1.35v QPI. The one that did not I got used from an EVGA forum member, and I'm pretty sure he degraded the chip.

[zoson]

Got some holiday cash and picked up a i5-2500K + Z68XP-UD4. I was surprised at how little there is to tweak . Currently running 5.0GHz as my daily overclock, here's the screenshot:

Settings:
100.2 x 50
Normal Vcore + 0.100V (~1.45 in BIOS, 1.404V under load in CPU-Z)
VTT 1.050V
CPU PLL 1.7V
CPU Overvoltage enabled
Turbo Disabled

Specs in sig

You sir, need to update your linpack binaries badly. You should be flying along at 130GFlops+, you're barely half that at around 70...

[l0ud_sil3nc3]

Got some holiday cash and picked up a i5-2500K + Z68XP-UD4. I was surprised at how little there is to tweak . Currently running 5.0GHz as my daily overclock, here's the screenshot:

Settings:
100.2 x 50
Normal Vcore + 0.100V (~1.45 in BIOS, 1.404V under load in CPU-Z)
VTT 1.050V
CPU PLL 1.7V
CPU Overvoltage enabled
Turbo Disabled

Specs in sig

and update to SP1 so you can fully utilize the AVX instruction set

[LennyRhys]

Also, I went nuts and was running my uclk at 4067MHz, which requires me to set vQPI to 1.49375v in bios. :O Luckily, my chip seems to be no worse for wear.

Does your chip not do 3600 uncore with 1.35v QPI? Optimally you want your uclk to be 2x your memory speed. So 1800MHz mem and 3600MHz uclk goes like bread and butter. Bringing your uclk up to 2x mem will also probably net you ~2 GFlops

Four out of the five gulftowns I've owned did 3600 uncore with 1.35v QPI. The one that did not I got used from an EVGA forum member, and I'm pretty sure he degraded the chip.

Even with HT off, 100+ GFlops at a shade over 4.5GHz is very good. And yes I noticed the ridiculous uncore freq and voltage you had - so far the highest vQPI I've used is 1.45, max CPU temp 84C, but I don't want to do that in a hurry again. I'll try my uncore at 3600 and see how I get on at 1.35v - no idea why I set it at 3200...I think I'm just too tired to be doing something as precise as this lol. In any case this 980X is a good one and might very well do 4.5GHz at 1.4v which is nothing to sniff at. I do have a lot of thermal headroom thanks to the very best air cooling and some very cold ambient temps (0C and below), so I'm definitely going to tweak 4.5GHz some more, and when I get an idea of temps/volts I'll decide whether a 4.6GHz run is on the cards.

Downloading CPU tweaker now and will get back on it either tonight or tomorrow. BTW my uncore was always at 3600 lol.

OK, I spent about an hour just watching how the tRL responded to changes in CPU, QPI and IOH voltages, and it was VERY interesting - thanks zoson for the spot-on diagnosis. Most often tRLs were set to 57,59,61 but at one point they tightened up to 58,59,60 which gave me an extra 2 GFlops in LinX, but still only 82 peak... I tried 56,57,58 for the heck of it and GFlops were the same... so I'm wondering what they would have been for GFlops to go as high as 84.5? Or perhaps it's Windows pagefile that's causing the irregularities? Still have to disable it.

I've found that changing vIOH is what sometimes causes the PC to boot with 4GB of memory detected, so I'm going to run memtest overnight tonight just to make sure the IMC and RAM are alright. There's a slim chance I might get a solid gold 990X soon (how golden? 4.5GHz on ~1.3v) in which case I'll definitely be hanging back on the voltages from now on. Once I'm sure IMC and RAM are ok, I'll get back on 4.5GHz and tweaking the memory subs.

[zoson]

Finally hit 4.7ghz. 1.6v :O
zoson | Core i7 980X [6C 6T] @ 4703.0MHz | 1.6v | ASUS Rampage III Extreme | Water | 3013A540

[LennyRhys]

Nice work zoson... that almost qualifies as a suicide run with that vcore lol.

Yesterday I wondered if my IMC and RAM were still ok, so I ran memtest for 7 hours overnight - no faults - and I ran LinX at stock for 2 hours this morning - also no faults, and the most insanely stable list of GFlops I've ever seen (63.4 GFlops every pass )

I've since been trying for hours to improve on my usual 83-84 GFlops at 4.5GHz HT and no matter what I do my rig just won't co-operate, even when I have control over all the variables. My memory subtimings are exactly the same every boot, and I've set tRL to help mitigate further inconsistencies.

Despite my efforts, the run that gave me 82 GFlops yesterday (standard boot, no tweaks) is giving me 58 GFlops with the exact same voltages and memory timings, right down to tRL being set to 58,59,60. I really don't understand what has changed that lost me nearly 25 GFlops, but the inconsistencies are really annoying me now seeing that others (Cryptik and McDown to name two) seem to get 86-89 GFlops with ease, and I simply cannot get to 85... Perhaps this 980X just has a sucky IMC...

180x25 for 4500MHz
CPU 1.4125v
QPI 1.35v
IOH 1.22v
RAM 1.65v

[R37ribution]

Nice work zoson... that almost qualifies as a suicide run with that vcore lol.

Yesterday I wondered if my IMC and RAM were still ok, so I ran memtest for 7 hours overnight - no faults - and I ran LinX at stock for 2 hours this morning - also no faults, and the most insanely stable list of GFlops I've ever seen (63.4 GFlops every pass )

I've since been trying for hours to improve on my usual 83-84 GFlops at 4.5GHz HT and no matter what I do my rig just won't co-operate, even when I have control over all the variables. My memory subtimings are exactly the same every boot, and I've set tRL to help mitigate further inconsistencies.

Despite my efforts, the run that gave me 82 GFlops yesterday (standard boot, no tweaks) is giving me 58 GFlops with the exact same voltages and memory timings, right down to tRL being set to 58,59,60. I really don't understand what has changed that lost me nearly 25 GFlops, but the inconsistencies are really annoying me now seeing that others (Cryptik and McDown to name two) seem to get 86-89 GFlops with ease, and I simply cannot get to 85... Perhaps this 980X just has a sucky IMC...

180x25 for 4500MHz
CPU 1.4125v
QPI 1.35v
IOH 1.22v
RAM 1.65v

First off your RTL is way to high for an uncore of 3600 and memory @ 1800 (with CL @ 7, if CL was @ 8 these settings would be fine) - channel A should be some where around 54-55 and then add 1-2 for each add'l channel. Also your tRAS at 24 with 7-8-7 timings seems high too, unless this is what your RAM is rated at. try something like 20-22. A LOT of things can affect your gflops, including CPU PLL, memory timings, not enough vcore/vtt. Its pretty crazy, but getting your timings correct will help a lot. I'm not even part of this club yet, but I'm almost stable at 225x17. Once I get that dialed in I'll try for a 230bclk and then up the multiplier on my 3520.

[zoson]

First off your RTL is way to high for an uncore of 3600 and memory @ 1800 (with CL @ 7, if CL was @ 8 these settings would be fine) - channel A should be some where around 54-55 and then add 1-2 for each add'l channel. Also your tRAS at 24 with 7-8-7 timings seems high too, unless this is what your RAM is rated at. try something like 20-22. A LOT of things can affect your gflops, including CPU PLL, memory timings, not enough vcore/vtt. Its pretty crazy, but getting your timings correct will help a lot. I'm not even part of this club yet, but I'm almost stable at 225x17. Once I get that dialed in I'll try for a 230bclk and then up the multiplier on my 3520.

:edit: I just noticed your vtt is only at 1.35v. Try going up to something like 1.45 or higher and rerunning the bench.

I've been trying to be nicer and more supportive... But I see a reply like this and I just *have* to chime in that nobody should listen to this. All of the above info is just flat out _bad_ and _wrong_.

RTL values will be different from board to board, chip to chip, and ram to ram. Unless you have the exact same memory kit, board, and chip from the same time period and bins, these values are not translatable from user to user.

Allow the board to set your RTL automatically, write down the values, and then simply tune them manually increment by increment to see if you can actually tighten them down more.

And anything over 1.35v VTT is _DANGEROUS_ for Gulftown processors.

[Mr. K6]

You sir, need to update your linpack binaries badly. You should be flying along at 130GFlops+, you're barely half that at around 70...

Forgive my noobishness, but is that as easy as getting them from Intel?

and update to SP1 so you can fully utilize the AVX instruction set

Thanks, just did

[R37ribution]

I've been trying to be nicer and more supportive... But I see a reply like this and I just *have* to chime in that nobody should listen to this. All of the above info is just flat out _bad_ and _wrong_.

RTL values will be different from board to board, chip to chip, and ram to ram. Unless you have the exact same memory kit, board, and chip from the same time period and bins, these values are not translatable from user to user.

Allow the board to set your RTL automatically, write down the values, and then simply tune them manually increment by increment to see if you can actually tighten them down more.

And anything over 1.35v VTT is _DANGEROUS_ for Gulftown processors.

The RTL settings have been known to auto detect too aggressively, you would know this if you browsed i4memory.com. Here is an article on it. There is a formula for RTL and also a link in that thread to an Excel sheet which can be downloaded to guesstimate your RTL. The value you get from the Excel sheet is usually within 1-2. You are correct, each board is different as the traces are longer/shorter going from the processor to the memory but its only going to throw off the RTL by a 1-2 usually and the Excel sheet as a parameter where you can set the latency from processor to RAM (it is defaulted to that of the R3E IIRC).

I apologize about the bad info on the vtt, I don't own a Gulftown. I do know that when my vtt is set to low my gflops are adversely affected on my 3520.

[LennyRhys]

It's not RTL causing my low GFlops - I've had them set at 58,59,60 for a while now and the GFlop fluctuation persists, but it's definitely more pronounced as I change voltages. I can lose as many as 10 GFlops with one 60mV increment...

After the above run with 58GFlops I re-ran LinX about 10 mins later after leaving my PC powered off. The test was run again with the exact same settings, and voila my GFlops have gone back to 84... but this time I got an error at loop 16. I can tell when GFlops are going to be high because the load temps are much higher, but like I said I didn't change any voltages or memory subtimings... this run is with page file enabled and standard boot, which has given me the best results yet.



As for fine-tuning tRL values - I tried, and it makes no difference. 56,57,58 yield the same results as 58,59,60; and anything below 56 on A doesn't POST.

As for VTT/QPI being too low, LinX still pulls 80+ GFlops below 1.35v, so I guess I just have to find the right ratio as Pilsy said. I've taken QPI all the way to 1.45v only to have diminished returns, so there's still LOTs of tweaking required here in the lower voltage register... I've never found overclocking such a challenge (or so frustrating) in all my life - nothing makes sense!!

[acebmxer]

Forgive my noobishness, but is that as easy as getting them from Intel?

Link is on the first page of this thread.

[Ace.]

very hot zoson

[zoson]

Forgive my noobishness, but is that as easy as getting them from Intel?

Thanks, just did

Get them directly from intel and extract the linpack_xeon32.exe and linpack_xeon64.exe's do you linx folder.

Link is on the first page of this thread.

The first page is no longer up to date.

The RTL settings have been known to auto detect too aggressively, you would know this if you browsed i4memory.com. Here is an article on it. There is a formula for RTL and also a link in that thread to an Excel sheet which can be downloaded to guesstimate your RTL. The value you get from the Excel sheet is usually within 1-2. You are correct, each board is different as the traces are longer/shorter going from the processor to the memory but its only going to throw off the RTL by a 1-2 usually and the Excel sheet as a parameter where you can set the latency from processor to RAM (it is defaulted to that of the R3E IIRC).

I apologize about the bad info on the vtt, I don't own a Gulftown. I do know that when my vtt is set to low my gflops are adversely affected on my 3520.

So, you told him to tighten his RTL values because auto detect is too aggressive? Way to change your story buddy.
Initially you said his RTL's were 'way too high' and to drop them to 54/55... Which is it?
Also, he doesn't have an R3E, he has a P6X58D.

It's not RTL causing my low GFlops - I've had them set at 58,59,60 for a while now and the GFlop fluctuation persists, but it's definitely more pronounced as I change voltages. I can lose as many as 10 GFlops with one 60mV increment...

After the above run with 58GFlops I re-ran LinX about 10 mins later after leaving my PC powered off. The test was run again with the exact same settings, and voila my GFlops have gone back to 84... but this time I got an error at loop 16. I can tell when GFlops are going to be high because the load temps are much higher, but like I said I didn't change any voltages or memory subtimings... this run is with page file enabled and standard boot, which has given me the best results yet.



As for fine-tuning tRL values - I tried, and it makes no difference. 56,57,58 yield the same results as 58,59,60; and anything below 56 on A doesn't POST.

As for VTT/QPI being too low, LinX still pulls 80+ GFlops below 1.35v, so I guess I just have to find the right ratio as Pilsy said. I've taken QPI all the way to 1.45v only to have diminished returns, so there's still LOTs of tweaking required here in the lower voltage register... I've never found overclocking such a challenge (or so frustrating) in all my life - nothing makes sense!!

RTL and tRL are different values. tRL is NOT exposed to the BIOS on any motherboard EXCEPT the R3E, and even then, it's ONLY with the 0003 beta bios, which is now no longer useful since it's got the smaller intel orom, vs 1502 which you can update the orom.

I suspect you're right and you're running into page file issues. Disable it and try booting into diagnostic mode. With 6GB installed and my page file off, I still can test 5120MB LinX. If LinX memory is paged out to disk... You will have a sharp drop off in GFlops.

I also suspect your core voltage is too low. My cpu will get into windows at 4585mhz with as low as 1.4v, but isn't stable until 1.4875v.
Also, I haven't seen any IOH's that needed more than 1.16v. The IOH doesn't have any impact on memory, so I'm not sure why that would make any difference. I bet you run into dropped memory channels again. Maybe your IMC is drifting enough to cause your hard set RTL's to be too low. You mentioned earlier that your c channel was detected at 61 cycles... but you have it set to 60.

Lastly, to keep every bit of performance, make sure you leave c-states enabled.

very hot zoson

My avatar, or my cpu?

[LennyRhys]

Oh... didn't realise tRL was different from RTL... meh. I'll put RTL back to auto and start from scratch with a lower clock speed like 4.3 or 4.4 just to give myself more thermal headroom and keep the voltages down, as I should definitely be able to get 80+ GFlops with 4.3GHz. The reason I set 58,59,60 is because that is what auto set on one of the best runs, however another good run was 58,61,64 (lol).

As for needing more voltage I'm honestly not sure - I've tried everything up to 1.5v on the vcore and 1.45 on QPI and the GFlops fluctuated so bad that it is almost impossible to tell - I've had 83-84 GFlops at both 1.4 vcore and 1.45 vcore... north of 1.45, GFlops sail closer to 75 and below.

[LennyRhys]

I gave up on HT - it's just a waste of time. Thought I'd see if I could get 100 GFlops so I turned HT off for the first time since getting this 980X and it took 5 mins of tweaking to get 101 GFlops... burning away just now and hopefully I'll have a new submission shortly.

OK got a couple of BSODs (once on loop 19 ) but I was pushing the uncore and memory a lot harder than I usually do to get the best performance without overdoing it on the voltages, and in the end I had to go over 1.4v QPI and 1.65v RAM but temps were very good, approx 80C above ambient. All in all this was INFINITELY easier than trying to get realistic GFlops with HT on - with only six threads the behaviour of the system is as you would expect: more voltage gives more stability; higher clocks give higher performance (I worked up from 200x23).

205x23: 4709MHz (CPUZ)
Vcore 1.48175v
QPI 1.4125v
RAM 1.68v
IOH 1.16v

LinX running:



And CPU Tweaker after stopping the run:



Zoson - thanks for all your help and patience with me, and for recruiting me to the GFlop police force LOL. Despite all my efforts and many long (and cold!) hours of laborious tweaking, 4.5GHz with HT on simply refused to give me consistent performance or behaviour in LinX, even in diagnostic mode with page file disabled. At least I now understand how LinX GFlops relate to real stability.

[Ace.]

your CPU zoson and i need a chiller for mine over 5Ghz with HT

[zoson]

4.7ghz with 1.4875v!!! That's really, really nice!!! Way better than my chip for sure!!! :O I went through 3x i7 970's and 2x i7 980x's to get the chip I currently have.
Congrats, and I'm sure if you work at it you'll figure out the HT GFlops thing. Looks like you'll be a good HPC thread member... If I ever get that started... I have to mine results out of here and figure out a way to make the tables more maintainable...

your CPU zoson and i need a chiller for mine over 5Ghz with HT

Yeah my chip runs cool, but needs LOTS of voltage. I did that 4.7ghz run at 25C ambient!!! I wonder if my chip would do better if I could get 0C temps like Lenny?

[LennyRhys]

You would probably get increased stability with lower temps, especially with a capable custom WC loop like yours. I managed to achieve exactly 80C above ambient (a steady 3C outside temp today) but I had to use a lapped copper TRUE and a 5500rpm/220CFM Nidec server fan to achieve it

This is with the Delta 4000rpm before I decided it wasn't man enough for the job: