M3A overclocking

[justapost]

The NB multi can be changed with wprcedit in the range ht-multi <= nb-multi <= cpu-multi.
Changing the ht multi via wprcedit results in no start after reboot.
Here is the description how to use wprcedit.
As it was mentioned before in other threads, go to Dev#24 Func#3 and change the bits 0-5 of the 32bit register D4, click Set and reboot.
NB-multi = D4:[0-5] + 4

Running sp 1M now with different HT/NB multis will post results later.

NB multis > 9 can not be applied with my 9500 MSR register c0010071h keeps the MaxNbFid in bits 63:59. In my case 00101b = 5 means 1800MHz max.
Using higher multis results in a lower ref HT as on other boards.

[Thesavage]

Yes you are correct, however not all applicatios exhibit the same behaviour switching between ganged and unganged.eg (below is a good analysis of the Phenom cpumemory subsystem)
http://www.lostcircuits.com/cpu/amd_phenom/9.shtml
Also the effect of northbridge frequency,
http://www.lostcircuits.com/cpu/amd_phenom/21.shtml
http://www.lostcircuits.com/cpu/amd_phenom/22.shtml
Also George ou posted a table with the link between memory speed core speed and the efficiency of the memory controller at various frequencies(if i could only find where i saved it)
And this would help to explain some differences at various frequencies.

[justapost]

Yes you are correct, however not all applicatios exhibit the same behaviour switching between ganged and unganged.eg (below is a good analysis of the Phenom cpumemory subsystem)
http://www.lostcircuits.com/cpu/amd_phenom/9.shtml
Also the effect of northbridge frequency,
http://www.lostcircuits.com/cpu/amd_phenom/21.shtml
http://www.lostcircuits.com/cpu/amd_phenom/22.shtml
Also George ou posted a table with the link between memory speed core speed and the efficiency of the memory controller at various frequencies(if i could only find where i saved it)
And this would help to explain some differences at various frequencies.

That you for the links. That article covers most of the current phenom issues.

I tried to find out why the results drop if i set a nb speed>9. Thought the bios adjusts the ref HT in a way that the nb speed does not exceed the speed at 9x. KTE mentioned that this is not the case. I ran CB10 4xCPU at 8 ht multi and 8-10 nb multi.
Difference between 8 and 9 multi ~200 (7100 vs. 7300). Between 9 and 10 multi the points firts dropped ~100 points.
I compared memory subtimings and found that the max async latency has changed.
Using the same latency the difference shrunked to ~30. I ran each bechmark five times and used the average.
Sems the drops where caused by timing changes.
Looked for a way to change the CPU VDDC via registers. Found the msr-register where the value is stored, but it is empty. I assume it is not used. Hope the next bios will allow changing the prozessor voltage again.

[KTE]

Achim, this is the difference between a real 140MHz difference in NB (8x vs 9x @ CPU 2640MHz).

-

You can confirm and cross check this with what macci reports.

Sorry, I wrote a lengthy post but Fx crash and I lost it. New build is extremely buggy.

[Thesavage]

If i remember correctly the memory controller is~
1) 93&#37; efficient @ 2200MHz
2) 96% efficient @ 2300Mhz
3) 100% efficient @ 2400MHz
This seems to tie in with the strange performance when overclocking. ie
Overclocked 2300 ~ 2600 can be slower than just a 2600.

[KTE]

Thesavage: Can you explain this in more detail please.
Also how do you know this?

I've not seen those figures of IMC speed reached apart from macci with ES and a BE chip. In all cases, higher IMC speed was quicker than lower. Max I've been is around 2250MHz with 9500 very unstable on stock NB 1.1VID. Max on both my 9600s was 1.98Ghz and 2GHz.

The last line of your post; are you referring to CPU speed or IMC speed?

[Thesavage]

Hi KTE
There was reference to this on a blog by George Ou he produced a table with the various clock speeds of Barcelona/ Phenom and the relative efficiency of the memory controller at that speed. (I saved it but may have deleted it with my last format) I will search further for it, but this was also raised as a topic on the old aceshardware. From that table the best memory eff was at 2.4GHz and at 2.8Ghz.
Maybe pm George Ou from Zdnet may be worth looking into. I was referring to the cpu speed comparisons which sometimes produce strange results, also there is a section i think at lostcircuits on the merits of the nb overclock.

[KTE]

Ah, ok, thanks.

I'll try and search for it to see what is around, but G.Ou has been very anti-AMD and thrown very emotionally hasty journalism since Barcelona was announced, so I don't think he's someone I would trust solely unless he cites credible sources and there is someone else who has also mentioned it with a decent reputation. I mean 100s of review sites, reviewers, Sami, AMD themselves, my own contacts, this forum members and their contacts including those working for DFI, OCZ, EVGA, MSI and Corsair, and none has mentioned this yet. I've seen it mentioned in no documentation either, so I'm a little skeptic here.

I'll try and see what I can find about it. It has to make sense architecturally to be believable though and I understand K10h uarch pretty well.

[Thesavage]

Got it KTE,
http://aceshardware.freeforums.org/v...114a745dc1c3db
2nd from the bottom.

[justapost]

Achim, this is the difference between a real 140MHz difference in NB (8x vs 9x @ CPU 2640MHz).

-

You can confirm and cross check this with what macci reports.

Sorry, I wrote a lengthy post but Fx crash and I lost it. New build is extremely buggy.

Great, thank you.

Here are my results, everthing at stock beside HT/NB/MasAsnycLtency

8xHT, 9xNB, 57ns MaxAsyncLatency

Mem Read/Latency: 6971 MB/s 67,6ns
L3 Read/Rite/Copy: 7228 MB/s / 7698 MB/s / 9495 MB/s

8xHT, 8xNB, 57ns MaxAsyncLatency

Mem Read/Latency: 6186 MB/s 77ns
L3 Read/Rite/Copy: 6404 MB/s / 6864 MB/s / 7946 MB/s

8xHT, 8xNB, 48ns MaxAsyncLatency

Mem Read/Latency: 6534 MB/s 71,1ns
L3 Read/Rite/Copy: 6700 MB/s / 6856 MB/s / 7942 MB/s

So lower multis work on M3A

[BrowncoatGR]

I thought memory on the K10 is derived from the NB clock not the CPU clock?

[KTE]

Got it KTE,
http://aceshardware.freeforums.org/v...114a745dc1c3db
2nd from the bottom.

OK. That clears things out.

I thought memory on the K10 is derived from the NB clock not the CPU clock?

Yep.

For reference: Those figures do not apply to K10 Phenom as it works out the clocks unlike K8. Those clocks Gearge posted about doesn't apply to Phenom. You can run memory dividers through the BIOS and get full 800 or 1066 speeds with Phenom. CPU/NB speeds are worked out like this:

CPU Speed = (1/2*HT ref.) * (FID+16) / DID
NB Speed = HT ref. * (FID+4) / DID

Memory, I'm not sure what is happening internally but on the outset, this is how it'll work:

Mem Speed = HT ref. * DRAM Divider

As far as I'm aware.

So lower multis work on M3A

Very nice Achim. Lower NB FID works on all boards as far as I've seen. Good to know you have it working too!

I would try Unganged mode instead, it'll give you better bandwidth and lower latencies.

[Thesavage]

Ok this is what i found overclocking just slightly @ different memory speeds 800vs667.
1 @ 800 raise fsb and memory speed raises with fsb.
2 @ 667 raise fsb and memory stays static?
Is this then correct.
His table does mention Barcelona./ and not Phenom although it should be comparible.

[KTE]

2 @ 667 raise fsb and memory stays static?

Are you sure?
Raise HT. ref to 210 and what RAM speed does it give you at 1:1.6 divider?

TBH, it's quite impossible to keep the RAM speed static when increasing HT ref. since HT ref. and the RAM speed is directly linked with the ratio you run it at.

It should be the same with any DRAM divider which works.

[Thesavage]

Yep ( raised fsb from 200-220 )thats why i have been slightly confused by the results out of this m3a board Im not to sure if this the same as other boards. I also found overclocking @ 667 is a breeze compared to 800 (although I am running 4x 1 gig) 667 seems to allow higher clocks as well as the ability to actually save and cold boot with those settings. May well be that aod is reporting wrongly will recheck with cpuz.

[justapost]

OK. That clears things out.

Yep.

For reference: Those figures do not apply to K10 Phenom as it works out the clocks unlike K8. Those clocks Gearge posted about doesn't apply to Phenom. You can run memory dividers through the BIOS and get full 800 or 1066 speeds with Phenom. CPU/NB speeds are worked out like this:

CPU Speed = (1/2*HT ref.) * (FID+16) / DID
NB Speed = HT ref. * (FID+4) / DID

Memory, I'm not sure what is happening internally but on the outset, this is how it'll work:

Mem Speed = HT ref. * DRAM Divider

As far as I'm aware.

Yeah, that's how it works as far as i know.
Maybe george had AM2 or single power plane opetron boards in mind.

Very nice Achim. Lower NB FID works on all boards as far as I've seen. Good to know you have it working too!

Good to see it's working.

I would try Unganged mode instead, it'll give you better bandwidth and lower latencies.

No bios option for unganged mode atm. Found Register F2x110h has a bit for ganged/unganged settings. WPRCEDIT shows only values <=FFh, so i can not tweak it that way.
The bios lacks support for real prozessor (VDD) and northbridge voltage (VDDNB) settings, tlb-fix and dc-mode.

[Jakalwarrior]

Have the ram settings available for AM2 chips changed in the newer bios releases? And uh... anyone heard of any board failures yet?

[justapost]

Have the ram settings available for AM2 chips changed in the newer bios releases?

Yep

And uh... anyone heard of any board failures yet?

U mean killed by ocing?

[Jakalwarrior]

U mean killed by ocing?

Killed by "normal OCing". Stuff you can do in the bios. Or just plain dead due to manufactuering defects etc... (modding or extreme cooling abuse excluded)

I bought the MSI budget 790X board before letting others be the guinea pig. My pioneering spirit died with the mosfets.

Also, how stable is the vcore? It doesnt appear to have as much power regulation hardware. Is it 4 stage?

*edit* found this on the asus page.
"4+1 Phase Power Design
Longer Life, & Higher Efficiency!
Unleashes ultimate memory performances with independent power to core components, while providing fast transient response and stability for the CPU under heavy loading or overclocking modes."

Sounds promising.

[justapost]

Yep ( raised fsb from 200-220 )thats why i have been slightly confused by the results out of this m3a board Im not to sure if this the same as other boards. I also found overclocking @ 667 is a breeze compared to 800 (although I am running 4x 1 gig) 667 seems to allow higher clocks as well as the ability to actually save and cold boot with those settings. May well be that aod is reporting wrongly will recheck with cpuz.

Damn, had to reinstall windows, suddenly it allway froze after autologin.
Checked memspeed oc'ed to 220 for 667 mem. CPU-Z reported correct, did not try aod.
Could push ref HT to 235 with 11x CPU multi and ram @ 800. Far from prime95 stable but it verified and finished super pi 1m (tlb-fix disabled).
On the M2A-VM i reached ~30,03s with 5-5-5-15 timings.

Killed by "normal OCing". Stuff you can do in the bios. Or just plain dead due to manufactuering defects etc... (modding or extreme cooling abuse excluded)

I bought the MSI budget 790X board before letting others be the guinea pig. My pioneering spirit died with the mosfets.

Yeah, saw the picture u posted yesterday. Played with the idea of buying one, rethinking that now.
I added all M3A related resources to the first post, none said something about defects.
But the bios is in a very early stage and does not allow high voltage increases, that makes it hard to reach the hw's limit atm.

Also, how stable is the vcore? It doesnt appear to have as much power regulation hardware. Is it 4 stage?

*edit* found this on the asus page.
"4+1 Phase Power Design
Longer Life, & Higher Efficiency!
Unleashes ultimate memory performances with independent power to core components, while providing fast transient response and stability for the CPU under heavy loading or overclocking modes."

Sounds promising.

790FX boards use 8+2 fyi.

[justapost]

Found another thread about the mobo at techpowerup here.

Interesting recommendations about nb temps.

[Thesavage]

Hey justapost just about to reinstall windows myself again, seems that after a bit of overclocking and testing windows freezes just after boot.(can usually hear the xfi complain) Asus should supply a free xp pro corporate with this board, give overclocking @ 667 a try seems slightly better on my setup. Yes where i live temps are very high in summer so i have noticed the nb temps are very hot, but the board remains stable.

[Ganki]

would you guys recommend upgrading the bios to the beta or stay with original until they release a new one?

with this bios i have been able to get to 3.0ghz out of my 4400+ on stock hs/fan
http://valid.x86-secret.com/show_oc.php?id=289736
i didn't run it to long as temps started to get up into the 60's
i been pretty lost on what ht divider to choose while o/c so i just set it to lowest while testing.

also i had the lan driver issue, i disabled the lan in bios and installed a different one, used it for a day, then when i had a o/c failure the bios got reset, i boot into windows and its been working even since. Hope i wont have to reinstall for a while, i decided to remove the extra card.

[Jakalwarrior]

Hey justapost just about to reinstall windows myself again, seems that after a bit of overclocking and testing windows freezes just after boot.(can usually hear the xfi complain) Asus should supply a free xp pro corporate with this board, give overclocking @ 667 a try seems slightly better on my setup. Yes where i live temps are very high in summer so i have noticed the nb temps are very hot, but the board remains stable.

Unless the PCIE etc.. is coming unlocked its usually ram overclocking that kills my windows installs get all of the cpu overclocking out of the way then use a memtest boot disk to test the ram. Saves the XP install.
Im sure yall already know that and are being lazy :P what can I say though, i do it too. More fun to get it into windows and benchmark it than test it for 20 minutes in memtest
Anywho, ill have one in 3 days. Ill ring her out and see if I can add anything to this.

[Thesavage]

Giving the testing a small break, and treating myself to some crysis and ut3 action.