Phenom II X4 955 overclocking adventure

[lopri]

Sylvanas: Please disregard my early 'theory' completely. I was reading AT's 955BE review and I don't think Everest can show memory 'lagging' behind L3. It looks rather like L3 can only process as much as it's fed by memory. I will look more into it as well as ACC. (BTW, have you tested negative ACC values? My board's BIOS defaults to -2 when left auto)

Oh and thank you for letting me know about AOD. I did not realize it gave so many controls. It's really cool that I can make profiles with core affinity. I used to do it manually in the past when I play games but it was pain in the a** to alt+tab out after game launch.

 

[Sylvanas]

@ Lopri: Thanks for the Bit-tech article, that explains alot- the voltages are a good guide too. I have not tested negative ACC values yet, and my board too defaults to -2% when on Auto aswell, I'll give it a go. Also, do you raise only the NB core for NB overclocking or other similar settings? I have 3 settings for NB voltages, NB core, NB HT and another I can't remember off the top of my head but I was under the assumption that anything other than NB core is applicable only if you are raising the HT above 2ghz and would not effect the overall NB speed? What voltages do you run?

@ Idontcare: I think you will find that the high OCs of 4ghz+ are all done on a 32bit OS and I can not remember seeing a 4ghz on high end air or water on 64bit so it's fair to say the OS is a limitation.

I have seen an improvement in the stability of 4ghz with a little more voltage (1.55) but it's not rock solid once so I've come back to 3.9 @ 1.48v. I have not seen any real improvement with ACC values above +2% and given that the Bit-Tech article only mentions 2% that may be all that's required. Going to test some lower multis and higher HTT today, see how it goes.

 

[lopri]

There were so many trial & error's so I cannot remember exactly but I believe I added a little bit of core voltage while tryng to raise NB. I was able to undervolt (1.312V) and overclock (3.6GHz) at the same time when NB was not touched. This is my BIOS setting.

Quite barren..

As you can see there is not much to see there. lol. And worse, all the voltages move by 0.1V except the CPU and CPU-NB voltages.. (ugh) It is one of the reasons why I settled at lowest voltages possible for most of them. There isn't really anything to glean from this BIOS setting.

 

[lopri]

Also, I am not sure if this is a norm for Phenom, but HT was not a factor/hindrance at all for my overclocking. It didn't seem to effect overclocking or performance so I left it auto or sync'ed it with NB. Naturally I didn't have to touch the voltage, but who knows what this BIOS does.. What is the default voltage for HT?

 

[Kuzi]

Got job Sylvanas. You know most overclockers can't get above 3.9GHz on Vista 64 with Phenom, there is some limiting factor. At least you reached the Max, and like you said if you try Vista 32 you will get +4Ghz for sure.

 

[geokilla]

Originally posted by: lopri
Also, I am not sure if this is a norm for Phenom, but HT was not a factor/hindrance at all for my overclocking. It didn't seem to effect overclocking or performance so I left it auto or sync'ed it with NB. Naturally I didn't have to touch the voltage, but who knows what this BIOS does.. What is the default voltage for HT?

HT (Hyper Transport) is shown to have very little to no effect on overclocking or performance.

 

[error8]

Just a bit off topic, but how on earth did you took that 4870X2 to 865 mhz, Sylvanas?

 

[Sylvanas]

Originally posted by: lopri
Also, I am not sure if this is a norm for Phenom, but HT was not a factor/hindrance at all for my overclocking. It didn't seem to effect overclocking or performance so I left it auto or sync'ed it with NB. Naturally I didn't have to touch the voltage, but who knows what this BIOS does.. What is the default voltage for HT?

It's 1.25v or something. I agree, I have not experienced any limitations related directly to HT frequency.

@ error8: It's under water aswell- full coverage waterblock which doesn't get much above the mid 30's under load (the GPU that is) the VRMs are kept cool aswell - it all helps in getting that core up .

Also have noted that -2% ACC does not affect stability/lower voltages much at all.

 

[lopri]

Can your sticks do 1066MHz? If so, try running 3:8. I swapped my sticks (P45 got DDR2-800 and DDR2-1000 to 955BE) and I saw immediate performance boost.

 

[Sylvanas]

Breakthrough for NB speed!

Wow it appears I missed a major option in my Bios. When entering the BIOS in the lower menu there is an option of 'change setup mode: press F9' by pressing F9 turns the BIOS to 'Advanced mode' just what I need! So after entering advanced mode a whole slew of settings are revealed which were otherwise not visible in 'Basic' mode- this applies especially to memory as there are many more subtimings than were available beforehand. Most interesting of which was when I was in the 'CPU feature' menu, an option appeared called 'CPU-NB VID' with a hex value of 24. I thought to myself "wtf? I already have that setting in the voltage menu" it appears that in the voltage menu there is a setting called 'NB-CORE' (which I have been adjusting all this time) and then this new setting in CPU Feature called 'NB VID' these are most definitely separate settings. So in fact all this time I have been adjusting the wrong setting hoping that it'll improve my NB speed. I think this 'NB CORE' may in fact be in relation to the 790FX chip on the motherboard itself and thus have no actual impact on NB at all, even though they share a similar name.

So, after finding a very helpful thread on this topic through 10seconds of google searching it appears that the default NB VID of a 955 CPU (at least on this board) is 1.1v! After identifying the HEX value that corresponds to a .1v increase in NB VID (1C) I dialled it in an hew presto I am now priming at 2.4ghz no problem- not quite done yet

Can your sticks do 1066MHz? If so, try running 3:8. I swapped my sticks (P45 got DDR2-800 and DDR2-1000 to 955BE) and I saw immediate performance boost.

In 1T mode this is a no go and I am yet to test them at 1066mhz in 2T if that will work however If it does I will have to compare with Everest as I remember an AT article a while back (and a few XS posts) that say that 1T 800 is more bandwidth than 2T 1066.

 

[Sylvanas]

In order to go above 2.2ghz NB fully stable, I needed to drop the CPU frequency 100mhz to 3.8- I guess to be more in 'sync' (I have seen a few exmaples of this reading around). Anyway fully stable at 2.6ghz NB 3.8ghz core OCCT 1hr here is a quick and dirty Everest run all was done at 1.35v NB VID 5-5-5-15 1T trc 22 TRRD 2 800mhz. Very pleased with this CPU and mobo good job AMD and DFI

 

[lopri]

Told yah. Your NB is now @2.6GHz! Congrats.

Now try lowering voltages by small amounts here and there for 24/7 setup!

WRT memory: On my board, 1T didn't do as well as higher memory frequency with 2T. DFI may be tuned for tighter timings, me thinks. But I agree, 1T won't be easily achieved @1066MHz. For me, it was easy to have two 2GB sticks running 1066MHz/5-5-5-15-2T, but 8GB has been a no go so far here. Even with the auto timing of 5-7-7.

 

[Sylvanas]

Originally posted by: lopri
Told yah. Your NB is now @2.6GHz! Congrats.

Now try lowering voltages by small amounts here and there for 24/7 setup!

WRT memory: On my board, 1T didn't do as well as higher memory frequency with 2T. DFI may be tuned for tighter timings, me thinks. But I agree, 1T won't be easily achieved @1066MHz. For me, it was easy to have two 2GB sticks running 1066MHz/5-5-5-15-2T, but 8GB has been a no go so far here. Even with the auto timing of 5-7-7.

Thanks I got there in the end. Next on the agenda is as you said, tweaking some voltages and some subtimings. thx for the input!

 

[Idontcare]

Originally posted by: Sylvanas
@ Idontcare: I think you will find that the high OCs of 4ghz+ are all done on a 32bit OS and I can not remember seeing a 4ghz on high end air or water on 64bit so it's fair to say the OS is a limitation.

I have seen an improvement in the stability of 4ghz with a little more voltage (1.55) but it's not rock solid once so I've come back to 3.9 @ 1.48v. I have not seen any real improvement with ACC values above +2% and given that the Bit-Tech article only mentions 2% that may be all that's required. Going to test some lower multis and higher HTT today, see how it goes.

Were all those LN2 demo's done on 32bit OS then?

I have a vapochill system sitting here unused (intentionally mothballed for future use)...if I were to get a 955 BE and put it under a vaporphase rig would I too not be able to get over 4GHz in Vista 64?

I'm interested because I had actually planned to do this at some point when I actually have the time to play again, but if there is some sort of 4GHz "barrier" then obviously I'm not about to waste my time and money getting snarled in that.

 

[lopri]

Originally posted by: Somebody

i have a problem
i bought a msi gd70 and a phenom 955 (0911BMPW) when the 955 came out
and for the life of me i cant get it to go over 3.9 i have tried higher fsb/cpu multi more volts (1.55) and nothing could it just be this chip is done @ 3.8?

Originally posted by: Gary Key

Hate to sound like a broken record around here, but what operating system are you using? If it is WinXP 64 or Vista 64, you reached the clock limit already on water.

:laugh:

 

[lopri]

Rofl there was another one.

Originally posted by: Gary Key

On average with air cooling, you will have about 200MHz more headroom with Vista 32 over Vista 64 at like settings once you hit the 3.7~3.8GHz range. On good water, the difference is about 300MHz between operating systems and on LN2, oh about 2.1GHz. Right around the 3.95GHz range with Vista 64 or Win7 64, you will hit a clocking wall. So far today, we finally hit 4.3GHz stable on V64 with LN2, exact same settings on V32 got us 6.4GHz.

He must be having fun.

 

[Idontcare]

Well that would definitely say 4GHz is not some hard-coded limit of the OS itself, unlike the timing loop issue with the original 400MHz K6-2 and Win95.

This just goes to show that the parts of the PhII chips needed to operate in 64bit situations have much less stability margin versus the parts needed to operate in 32bit OS'es.

If memory serves this was also noted for the original Phenom chips as well. 64bit overclocks were always less than 32bit OC's for the same chip.

Thanks for the quote Lopri, I must admit the datum regarding LN2 OC being only 4.3GHz kinda squashes my thoughts on putting my vaporphase gear to work on a PhII. Was more thinking I'd hit 5GHz territory...but 4GHz on a quad? I had that back in 2006. Nothing terribly exciting about doing it now. I guess I wait for Bulldozer then.

 

[lopri]

I thought it sounded like some kind of an errata.

 

[lopri]

I haven't been able to toy around my 955BE lately but I am still obsessed with Phenom's NB. I did notice some 'strap' like behavior, however. It seems to me that AMD tried to optimize the performance around certain NB frequencies, depending on CPUs. For example, I get better latency by booting @3.2GHz and clocking up to 3.8GHz in Windows using AOD, than by straight-booting @3.8GHz. Of course everything should be same since I use multipliers, but not. Intel's MCH used to work similarly so some overclockers would boot at lower FSB then clock up in Windows in order to take advantage of lower straps. Also of interest is that AMD has been making server CPUs with 2.2GHz NB for some time. (I didn't know about this until like yesterday)

Some other random observations:

- 1T under 3:8 divider (DDR2-1066) is a no-go. But DDR2-1066/1T is still possible by raising the HTT under 1:4 divider.
- Cool'n Quiet doesn't work if overclocking is done by multipliers. It's techinically understandable yet somewhat disappointing because it defeats the point of "Black Edition" I'm trying to find a sweet spot for native C'nQ without resorting to 3rd party utilities.
- This CPU is fast as hell. (duh) I dug up some of my old screenies (like X2 4800+, Opteron 146, etc.) and clock-for-clock improvement over K8 is significant. It's unfortunate AMD couldn't manufacture K10 at today's frequencies when launched..

There is a discussion going on WRT Phenom's NB @XS. Tony = the creator of famous Bigtoe BIOS for DFI NF4.

 

[Sylvanas]

Nice post Lopri, I have been keeping an eye on that XS thread aswell, very interesting. The strap behaviour you saw, what kind of gains are you noting? Is it a sinple matter of subtimings adjusted to NB frequency or is there some hidden variable possibly like Intels TrD. I have backed off the NB voltage by 1 notch and all is well and I am quite happy with the way things are going time to enjoy the new CPU instead of torture it . Just for interest I have tried 2.8ghz on the NB with even 1.4v and it was a no go and required a CMOS reset- perhaps there is a hard lock limit for NB on vista 64 like CPU speed 4ghz.

 

[lopri]

At the end of my 955BE thread, I did some quick cpuz latency tests with varing NB frequencies. I learned that as long as CPU and NB are at a same frequency, CPU always reaches L3 in 41 cycles. (at least according to cpuz latency) As the gap between NB frequency and CPU frequency widens, it takes more cycles for CPU to reach L3. (naturally) All the while, L1 and L2 latencies didn't budge. (L1 = 3 cycles, L2 = 15 cycles)

See, if I boot the system w/ 3800MHz CPU and 2200MHz NB, it will give me predictable value which are L1 = 3 cycles, L2 = 15 cycles, L3 = 55 cycles. If I boot CPU 3800MHz, NB 2600MHz, then I'll get L1 = 3 cycles, L2 = 15 cycles, L3 = 49 cycles. Now, if I boot the system w/ 3200MHz CPU and 2600MHz NB and increase the CPU frequency using multipliers in Windows -> See what happens

Incredibly, L1 latency is reduced to 2 cycles and L2 latency is reduced to 12 cycles! How is that possible?!

I read original Barcelona articles from it.AnandTech when I have free time, and there are a lot of fascinating information there. (such as this) Johan is such a genious and it seems like he figured a lot of stuff out like 2 years ago..

 

[poohbear]

btw what's the max vcore u guys are comfortable giving these 45nm cpus? i read anything beyond 1.45v is dangerous. U guys follow that or just watch temps?

 

[lopri]

AMD's guidline is 1.5V. But as you may have learned by now, what holds these chips is not the execution engine but NB/L3. In order to see how far your chip can clock, I suggest you underclock/undervolt NB. Try something like 1600MHz~1800MHz. (and undervolt if possible) Lock NB frequency there and just raise the core frequency using multipliers. You're likely to be surprised.

Also, apparently some older boards (especially nForce variants) have vCore and vNB tied toghether, and raising vCore will increase vCPU-NB as well. Under such a condition, I wouldn't touch the voltage. See if your board allows voltage adjustment for CPU-NB (note this is not the northbridge on the board, but memory controller in the CPU).

 

[soonerproud]

I tried to overclock to 3.6 by upping the multiplier by .5 increments at a time and the system was completely stable through OCCT/Linpack and 8 hours of Prime 95. If I shut the PC completely down then the system would not even boot to the bios, but works fine once I clear the CMOS using the button on the back and re-flash and configure the BIOS once again. If I reboot, it is fine but if I try to start the PC from a cold shut down, I have the same issue.

Is this a voltage issue, do you think? My BIOS forces me to manually set the northbridge multiplier when I use the CPU multiplier. Do you think it might need me to manually set the voltage on the northbridge? What would the voltage for the NB usually be at stock speeds?

I am perplexed how a system that is completely stable will not post on a cold start up. What do you think I am doing wrong here?

(Gawd I miss the old Socket A days where I just used DDR400 memory with my XP2500 and just upped the FSB to run it at a 3200! Over-Clocking is so much more difficult today.)

 

[geokilla]

Originally posted by: soonerproud
I tried to overclock to 3.6 by upping the multiplier by .5 increments at a time and the system was completely stable through OCCT/Linpack and 8 hours of Prime 95. If I shut the PC completely down then the system would not even boot to the bios, but works fine once I clear the CMOS using the button on the back and re-flash and configure the BIOS once again. If I reboot, it is fine but if I try to start the PC from a cold shut down, I have the same issue.

Is this a voltage issue, do you think? My BIOS forces me to manually set the northbridge multiplier when I use the CPU multiplier. Do you think it might need me to manually set the voltage on the northbridge? What would the voltage for the NB usually be at stock speeds?

I am perplexed how a system that is completely stable will not post on a cold start up. What do you think I am doing wrong here?

(Gawd I miss the old Socket A days where I just used DDR400 memory with my XP2500 and just upped the FSB to run it at a 3200! Over-Clocking is so much more difficult today.)

So you've tried updating the BIOS right? I'd say make sure everything in your BIOS is set to manual and maybe raise the voltages by a notch or 2...

Also why are you using Linpack? I thought Linpack worked for only Intel CPUs. Plus Linpack, imo, does nothing but generate excessive heat. I prefer OCCT over Linpack.

@lopri. AMD overclocking is SO difficult now. You should write up a guide and see if AnandTech will let you post it on the front page. This thread is very resourceful for Phenom II overclockers.