Nice reading, nice results :)
I whish I wouldn't have this strange 30s lag after running 2 WUs for a time. As Serious Sam once proudly said:

Double the gun, double the fun!

MrS

---

Scanning for our furry friends since Jan 2002

i've been running 2 WU's at a time for the past few days, and my increase in PPD seems to be more inline with Zydor's estimate of 5000...

i've decided to start testing from scratch. i got the above figure by simply looking at my BoincStats page which, as you know, isn't the most accurate way to do it. so i'll switch back to 1 WU at a time and keep an eye on elapsed crunch time. then i'll do 2 WU's at a time again and see how the times compare. i'll report back what i have some solid numbers...

---

Hello,
has anybody tried running 2 threads on a cayman GPU with lowered memory clocks? I've got a 6950 (unlocked to 6970, from Sapphire with default layout and 2GB).
I usually use 880MHz gpu and 500MHz memclock with 1,065V, works fine with only one thread. When running 2 threads, i get vpu recovers and sometimes freezes with that settings.
The problem is not caused by temperature (tried fanspeeds of 50%), not caused by the gpu clock (also tried only 800Mhz at 1,1V, went even lower), but is caused by the memclock. When running default memclock, mw runs fine even with two threads, but the power consumption is much higher.
The problem appeared with the cat. 11.1a, earlier drivers worked fine, even with 2 threads and lowered memclock. I already reported in the amd forum. So can anybody confirm my issue, or is this just my card/vga-bios?
greets.

I'm running 2 concurrent wu's on (3) 6970's @ 950 core 685 memory and 1200mV with no issues.
I did have to put:

<cmdline>--gpu-target-frequency 60</cmdline>

in my app-info to get rid of the screen lag

This was not necessary with the previous app
can also go all the way to 1Ghz on core but no need now as it's already no.1, LOL

Hello,
has anybody tried running 2 threads on a cayman GPU with lowered memory clocks? I've got a 6950 (unlocked to 6970, from Sapphire with default layout and 2GB).
I usually use 880MHz gpu and 500MHz memclock with 1,065V, works fine with only one thread. When running 2 threads, i get vpu recovers and sometimes freezes with that settings.
The problem is not caused by temperature (tried fanspeeds of 50%), not caused by the gpu clock (also tried only 800Mhz at 1,1V, went even lower), but is caused by the memclock. When running default memclock, mw runs fine even with two threads, but the power consumption is much higher.
The problem appeared with the cat. 11.1a, earlier drivers worked fine, even with 2 threads and lowered memclock. I already reported in the amd forum. So can anybody confirm my issue, or is this just my card/vga-bios?
greets.

Not a card or driver problem - pure laws of physics. The driver reset is because it needs more clout to cope with whats being thrown at it. When you went to two WUs, more is going up and down cpu-gpu. When you went to defaults re memory, you increased the bandwidth available. That resolves the problem as now more data can pass at anyone time cpu-gpu. (memclocks is a measure of bandwidth, the amount of data that can travel at anyone time, not its speed)

The higher bandwidth needed more power to run it, normal. Probably find you could reduce the memory to circa halfway between old setting and defaults, would need to play with it one step at a time.

As to the differences with old driver "which worked", no definitive answer other than drivers evolve, parts get rewritten, particularly at present with 6XXX, as the driver will have to cope with the ability to have separate kernels to run two separate applications without interference. The latter ability of 6XXX not yet really out in the open - but is part of the new GPU architecture design - suspect it will be more apparent as time goes on and software written to use its ability to run two separate applications without interfering with each other.

The "why" in the last para is slightly speculative, but I'll bet is pretty close as any re-write of the 6XXX drivers must take into account two kernals running, even though its not widely used yet.

Bottom line - I'd say its normal for the driver as is, just find the new memclock level needed and crunch on.

Regards
Zy

You might want to try Cat 11.3 or 11.4, if you haven't already done so. Mine works with 2 WUs at 900 MHz core (1.10 V) and 625 MHz memory. Didn't test any lower.

Physically nothing should change when running 2 WUs instead of 1, as only 1 WU is active at any one time anyway. The card is just switching quickly between them.

MrS

---

Scanning for our furry friends since Jan 2002

Hello,
i did some further testings and it seems, getting no freeze at default clocks was just luck.
I wanted to run default clocks and reduce the memclock step by step and even got a freeze at the default settings without changing anything at all (even after a fresh reboot).

So i can't run more than one task of MW at all. I'll stay at the settings i used before and keep running only one task. Maybe it has to do something with the shader unlock (i did the RBE unlock, didn't flash the 6970 bios to keep the low voltages).
As mentioned, the problem appears when using cat. 11.1a, 11.2, 11.4 preview and 11.4 (final).
Perhaps AMD find's a solution in upcoming drivers, i'll keep trying.

Another solution could be, to do the CPU work at the end of the WUs after starting a new WU on the GPU. This way there wouldn't be any need for running 2 WUs, as the GPU would have no free time at all. But i don't know if that's possible.
greets.

Last weekend I updated from Cat 11.3 to 11.4 and 1 WU runs fine. However, when I try 2 WUs in parallel my Win 7 65 Aero desktop freezes almost instantaneously. As I said before, with Cat 11.3 it took about 1 day of run time after which ~30s "freezes" / pauses would appear. The current behaviour is very different.

I've also got the shaders of my HD6950 unlocked, but don't use the HD6970 bios to keept the low voltage. MSI Afterburner could adjust the voltage by now, but this way it's easier.

MrS

---

Scanning for our furry friends since Jan 2002

Can the .62 use the GPU in the AMD chipsets like it could a few months ago?

You're probably mixing this with Collatz. MW has never been able to run on such GPUs due to the lack of double precision support.

MrS

---

Scanning for our furry friends since Jan 2002

@Zydor - I read with interest your earlier post to Oberst regarding memory speeds. I always understood that at Milkyway, it was best to lower the memory speeds as low as possible to save on power and heat. I've done that with my 5850 and lowered it to 500Mhz and its running a lot cooler.

But that is only running a single workunit. Are we now saying that if I wanted to run 2 workunits at a time I would probably have to up the memory speed to compensate? Or was your advice only meant for the 6000 series cards?

---

Don't drink water, that's the stuff that rusts pipes

@Zydor - I read with interest your earlier post to Oberst regarding memory speeds. I always understood that at Milkyway, it was best to lower the memory speeds as low as possible to save on power and heat. I've done that with my 5850 and lowered it to 500Mhz and its running a lot cooler.

But that is only running a single workunit. Are we now saying that if I wanted to run 2 workunits at a time I would probably have to up the memory speed to compensate? Or was your advice only meant for the 6000 series cards?

Not in a definitive "catch all" sense. I was suggesting that a slightly increased clocks is likely to be normal given two WUs going through at once.

*queue zillions of posts saying I didnt have to*

Point is, if there needed to be an increased memclock to get two working, its nothing to worry about, and any increase in heat would be normal if memclocks were increased.

Should it be needed (the increase), in theory no, as pointed out above the card will flip from one WU to another. It *might* make a difference if both were sent down the pipe at once start/end due to larger bandwidth. Probably unlikely as MW datasets are small.

However bottom line is, if in the real world, theory or no theory, individual finds it necessary to increase memclocks slightly, no worries, the subsequent increase in heat would be normal.

Pays your real world money, takes your real world choice, to abuse a classic phrase :)

Regards
Zy

Thanks for the reply Zy, appreciated :-)

---

Don't drink water, that's the stuff that rusts pipes

I am sorry but I am not deep enough into this to know what projects my cycles are processing. What I do know is that up until about a month or two ago both my video card and the HD3200 GPU in the chipset were processing tasks from Milkyway@home. In fact I was a little surprised that both the 4875 and the 3200 produced almost equivalent output. I had to disable the use of my chipset GPU with the recent revisions of the client (.58) after the server issues. I have heard/seen the double precision issues before but the ATI HD3200 seemed to have the capabilities before. If the hardware does or does not have the capabilities, I have not bothered to investigate. But with the multiple revisions of the client recently, being in the software industry, I thought it worth asking if the GPU is again usable. It would be sad to lose all the calculations of a GPU because of no communications.

MW has always required double precision capable GPU's.

Hello,
me again. I installed a brand new HD5830 (is now on discount as "Sapphire 5830 Extreme") in my server, to let him crunch mw too. After installing the drivers manually (seems like amd blocks the catalyst installation on server 2k8r2 since cat. 10.8) the card is up and running. So i tried running 2 wus again, the system freezed right after launching boinc. Seems like cat. 11.4 has some trouble with 2 mw wus at a time, not only limited to hd69 cards.
Greets Oberst.

Yeah.. the same being reported at SetiGermany. Cat 11.4 really dislikes running 2 WUs. With some older ones it's still "somewhat OK".

MrS

---

Scanning for our furry friends since Jan 2002