What's the current Llano APU vantage wr?

[The Stilt]

I've not run into any circumstances where passing 24 hours of memtest in windows with one instance per CPU core yields anything less than complete memory stability, even on this chip. What memtest are you using where that isn't the case? I was playing DNF for hours before I fired up memtest, so it seems stable enough for me. Perhaps you're closer to the ragged edge of stability than I am at these current clocks?

--Matt

I can do Memtest86 or HCI Memtest at 933MHz+ as long as I want to without any errors.
In 06 GT-1 the system crashes usually in the first pass.

This is IMC related, not the ram.

[El Gappo]

Hypers on the MSI, good news; they work, bad news; no they don't.

Performing like channel and bank interleaving are disabled but they aren't and it hasn't reverted to single channel. p882 in 11 @stock 1866 7-8-7 as opposed to p1100+ on the biostar with the exact same settings.
Running the same board with the gskills now, a little slower than the other boards cpc but not massively out of touch like with the hypers. It's also killed my debug card which I'm very not happy about.

Sorry chew, thread well and truly hijacked.

GIGABYTE A75-UD4H.

From the looks of this thread, it should be serving my overclocking needs quite well

uhoh

[mattkosem]

I can do Memtest86 or HCI Memtest at 933MHz+ as long as I want to without any errors.
In 06 GT-1 the system crashes usually in the first pass.

This is IMC related, not the ram.

That hasn't been my experience, with these sticks at any rate. I haven't pushed the ref clock past 120 with them yet though so the IMC is probably under much less durress than yours is in that scenario. Even with the flares at 1000+ though (their rated speed), if I can't pass the last 3 versions of 3dmark, I can't pass a full sweep with 4 instances of memtest. The other dead giveaway is that windows hangs briefly at a black screen when UAC prompts appear if Aero is on when the clocks are too high or timings are too tight.

I'm not sure what your stability testing regimen looks like, but I don't bother using clocks that won't pass an hour of 8k prime with one thread per CPU thread and all of the free ram in use and at least 200% in memtest with one instance per CPU thread. Perhaps those two catch the issues that cause 3dm06 to die.

--Matt

[chew*]

Matt I have to agree with Stilt on this one.

3d memory clocks "ram dependant" is much harder on Llano than anything you can run in 2d.

Certain IC's that pass 2d can't even pass 3d at stock..........

[mattkosem]

Matt I have to agree with Stilt on this one.

3d memory clocks "ram dependant" is much harder on Llano than anything you can run in 2d.

Certain IC's that pass 2d can't even pass 3d at stock..........

Interesting. Perhaps the ICs that I have aren't of the affected type(s)? I've had this since Launch day and haven't had any issues with 3D when 2D is "rock solid"(tm). I've only ever used it with 4GB dimms and haven't strayed too far from stock ram clocks/timings if that makes any difference. I also have Aero on at all times (except when 3dmark disables it) so I suppose there's always SOME level of 3D usage at all times, however light it may be.

--Matt

Edit: Hmm, well, it passes 10 runs of 3dmark06 at these settings. Perhaps these sticks have more in them than I thought. Orb seems to be down, so I can't submit my last score, but it's
rock solid at that as far as I can tell. I'm vetting out this setup (cpu, ram, mobo) before migrating my HTPC over to it. Any other good stability tests I should run?

Edit2: ORB is back up now. It's slower than the worst score that I have on my account (PHII 940/4830) by roughly half. I bought it for HTPC use, they can't all be winners. : http://3dmark.com/3dm06/15956127 And, as not to veer too far off topic, the vantage score is barely faster than stock clocks with faster memory and quite a bit slower than a mild OC with faster memory.

[crazydiamond]

chew out of curiousity have you tried hypers with dedicated gpu & disabling apu gpu?

[chew*]

chew out of curiousity have you tried hypers with dedicated gpu & disabling apu gpu?

No I imagine it would be fine however. The dedicated gpu uses it's own memory and not system memory.

[crazydiamond]

right, just wondering how they clock.. like thuban hypers or regular hypers

[chew*]

right, just wondering how they clock.. like thuban hypers or regular hypers

Frequency wise they seem to do better than thuban on this board, timing wise ( performance oriented ) I really haven't put to much time in other than to see if I could tune the 3d issues away on the APU.

[mattkosem]

I think I found the missing link between the results you guys are describing and what I'm seeing. I can even run 3dmark06 these sticks at 1600 C7 till the cows come home with the ref clock at 120. Once I push the ref clock up though, even with roughly similar ram and cpu core clocks, THEN I see what you all are hollering about. I'm able to start seeing failures in 3dmark but not anything cpu-related with these guys at around 1733 at C8 (130 ref clock). Interestingly, with my flares, this same exact clock/voltage configuration on the APU side with the ram at 2080 pass it over and over again. And, while the flares will run 150 on the ref clock these xms3 sticks won't, even at stock clocks and timings. Seems whatever ICs are on these sticks must behave differently with the IMC, as you mentioned might be the case.

This is sort of akin to increasing the NB clock on Deneb/Thuban, which also imposes a lower limit on clocks and timings. I'm guessing it's this way because the exact same thing is happening because the multis are all locked?

--Matt

[The Stilt]

Found the damn thing that messes with the memory on Asus boards.
Disabling bank interleaving does the trick.

Asus boards do not have an option to disable it by default.
And IIRC it cannot be turned off on the fly due memory training requirements.

I fiddled a bit with the bios and found the option included in all of the current bios versions.
There is a option for it but it is hidden and cannot be adjusted by user.

After disabling the bank interleave all of the 3D stability problems at 1600MHz+ MEMCLK went away.
Now I can easily do 2000MHz CL8 with these D9KPT sticks 3D stable.

Naturally disabling the BI will hurt memory performance (slightly), but the difference is nothing compared to the performance hit caused by lower MEMCLK or single channel mode.

I will post a "fixed" bios for F1A75-M Pro board tomorrow. Besides the BI disabling option, I updated the Sumo VBIOS code to the most recent version.
If necessary I can make the same fixes for all of the F1A75 family boards.

Attachment 118973

[TsunamiJuan]

Looking forward to trying this out Been conformaling boards most of the weekend. So should get some more crazy ln2 runs in soon.

[The Stilt]

I guess it is "tomorrow" already...

Asus F1A75 0307TS

Asus F1A75-M 0406TS

Asus F1A75-M LE 0308TS

Asus F1A75-M PRO 0802TS

Asus F1A75-V EVO 0603TS

Asus F1A75-V PRO 0902TS

All of them have the same changelog:

- Added option to disable memory bank interleaving.
- Updated Sumo VBIOS code from version 012.043.009.000.000000 to 012.043.014.000.000000

The option to disable or enable bank interleaving is located in "Advanced -> IMC Extra" page.

Do not disable it unless you have stability issues at high MEMCLK frequencies.

Attachment 118976

And here will be the usual disclaimer: No compensation, no support, no nothing. Use it and you are on your own.
In case something goes wrong you will either need a working motherboard with DIP8 socket or SPI programmer to recover.

Flash with EZ Flash 2 or AFUWIN (x86 or x64).

[chew*]

[Dumo]

Awesome

Thanks Stilt

[Alex-Ro]

Good job Stilt.Are there any other options hidden that can be revealed ? Something like Write cas latency,maximum read latency,idle cycle limit regarding mem timings?

[The Stilt]

Good job Stilt.Are there any other options hidden that can be revealed ? Something like Write cas latency,maximum read latency,idle cycle limit regarding mem timings?

There is plenty of "extra" settings available, however they are mostly for developer use only.
All of the memory related parameters are already in use, so the missing ones can only be adjusted by writing to the IMC registers directly.
Naturally timings like tCWL cannot be adjusted on the fly.

[chew*]

Newish Bios.

Flash at your own risk etc etc, not responsible for damages etc etc, you know the drill.

http://members.cox.net/wmdieselmc26/A75UD4H.F5g

Tested UD4H with Bank Interleaving disabled and D9KPT, anything under 1.72 failed.



Tested Hypers with Tightish timings but less speed conservative voltage ( 1.65 IIRC ) Bank Interleaving Enabled.

[El Gappo]

Schweet!

[mattkosem]

Newish Bios.

Flash at your own risk etc etc, not responsible for damages etc etc, you know the drill.

http://members.cox.net/wmdieselmc26/A75UD4H.F5g

Tested UD4H with Bank Interleaving disabled and D9KPT, anything under 1.72 failed.

Tested Hypers with Tightish timings but less speed conservative voltage ( 1.65 IIRC ) Bank Interleaving Enabled.

So 1000 cas8 at 3500 with interleaving off in XP is roughly what I get at 3500 at 1000 cas 9 (9-10-9-25-35) in 7 with it on - same bus speed same multi. 8657

What do you get in vantage at those settings? I see P5074.

--Matt

[chew*]

So 1000 cas8 at 3500 with interleaving off in XP is roughly what I get at 3500 at 1000 cas 9 (9-10-9-25-35) in 7 with it on - same bus speed same multi. 8657

What do you get in vantage at those settings? I see P5074.

--Matt

Ahh I posted that stuff up more or less just to show it passed, I wouldn't focus to much on the actual scores.

I was riding with very comfortable and lazy settings. The scores actually suck. I can look into maxxing them out later at those speeds for a better idea.

[chew*]

Ok matt ran some stuff off.

Timings used, quite a bit in scores can be gained by tuning.



I ran off 06 at 2 diff straps, obviously 150 ref clock will net you better results, also I'm not to fond of the high divider on the UD4H.

125 ref clock


150 ref clock


Xp 150 ref clock


I'd like to say I'm preety good at analyzing performance/effeciency

That said

These scores should be easy to beat on the ASUS with stilts bios and win7 64 w/ cat 11.7 with the same hardware, the only thing that looks weak on the ASUS board is firefly forest ( GT2 ) unless he had his gpu above 600, if so you can discard this.

For reference I think my Xp is using cat 11.7 and my win7 32 is using some cruddy 11.6B still.

I ran vantage off for you to.

125 ref clock


150 ref clock

[The Stilt]

The bios code on Asus F1A75 -series boards is quite unoptimized I must say...

The driver compensation and address/command registers are not configured properly.
Or actually... They are not configured at all

After programming these registers with the correct values I can now run benches with ~1080MHz MEMCLK.
When left to the default values Asus sets them, the maximum stable frequency was ~1020MHz.

(1600 or 1866MHz MEMCLK, DIMM present in both channels, S.A.M Disabled (i.e 1T))
These ARE the correct settings for all A-series APUs as long as the conditions described above are met.

Driver Compensation:

CKE Drive Strength (CkeDrvStren): 1.50X
CS/ODT Drive Strength (CsOdtDrvStren): 1.50X
Address/Command Drive Strength (AddrCmdDrvStren): 2.00X
MEMCLK Drive Strength (ClkDrvStren): 1.50X
Data Drive Strength (DataDrvStren): 1.25X
DQS Drive Strength (DqsDrvStren): 1.25X
Processor On-Die Termination (ProcOdt): 60ohm

Address / Command:

CKE Setup Time (CkeSetup): 1 MEMCLK
CKE Fine Delay (CkeFineDelay): 19/64 MEMCLK
CS/ODT Setup Time (CsOdtSetup): 1/2 MEMCLK
CS/ODT Fine Delay (CsOdtFineDelay): 0/64 MEMCLK
Address/Command Setup Time (AddrCmdSetup): 1/2 MEMCLK
Address/Command Fine Delay (AddrCmdFineDelay): 0/64 MEMCLK

Try it and let me know.
In case the memory itself is the limiting factor I doubt these settings make any difference
IMC troubles it should fix however.

[ZenEffect]

the stilt back in action! good to see you around again buddy

[mattkosem]

The bios code on Asus F1A75 -series boards is quite unoptimized I must say...

The driver compensation and address/command registers are not configured properly.
Or actually... They are not configured at all

After programming these registers with the correct values I can now run benches with ~1080MHz MEMCLK.
When left to the default values Asus sets them, the maximum stable frequency was ~1020MHz.

(1600 or 1866MHz MEMCLK, DIMM present in both channels, S.A.M Disabled (i.e 1T))
These ARE the correct settings for all A-series APUs as long as the conditions described above are met.

Driver Compensation:

CKE Drive Strength (CkeDrvStren): 1.50X
CS/ODT Drive Strength (CsOdtDrvStren): 1.50X
Address/Command Drive Strength (AddrCmdDrvStren): 2.00X
MEMCLK Drive Strength (ClkDrvStren): 1.50X
Data Drive Strength (DataDrvStren): 1.25X
DQS Drive Strength (DqsDrvStren): 1.25X
Processor On-Die Termination (ProcOdt): 60ohm

Address / Command:

CKE Setup Time (CkeSetup): 1 MEMCLK
CKE Fine Delay (CkeFineDelay): 19/64 MEMCLK
CS/ODT Setup Time (CsOdtSetup): 1/2 MEMCLK
CS/ODT Fine Delay (CsOdtFineDelay): 0/64 MEMCLK
Address/Command Setup Time (AddrCmdSetup): 1/2 MEMCLK
Address/Command Fine Delay (AddrCmdFineDelay): 0/64 MEMCLK

Try it and let me know.
In case the memory itself is the limiting factor I doubt these settings make any difference
IMC troubles it should fix however.

Wow, way to be. I can kick my base clock up to 130 and cpu down a multi without even touching any of the other timings after setting this, let alone mutilating them like I previously had to do to get the ram to run stably at 1040. Instead of losing points, now I gain 200 without adjusting anything else. P5265 Perhaps I'll have some time to try more tomorrow.

I really wish I could remember what I was running on my best score so far (P5523). I really need to buy the real version of vantage one of these days so I can screenshot this stuff.

--Matt