GPU WU's not competing before final deadline

[bebop]

I have an ATI 3450 chip on my GPU (ASUS $40 video card) and I'm running the F@H GPU client 6.23. I have been trying to complete a WU for about 2 weeks and haven't been able to finish one before the final deadline hits. I don't end up getting any points and would like to be able to finish a WU on my GPU under the time limit. Is there a way/setting to download WU's for the GPU that take less time to complete? Or perhaps a setting on my system that can boost performance? I tried over-clocking (ATI Catalyst Overdrive settings) the GPU Clock and the Memory clock - but that didn't do much of anything for completing the WU's on time.

------
Display info:
Name ATI Radeon HD 3400 Series
PNP Device ID PCI\VEN_1002&DEV_95C5&SUBSYS_01F41043&REV_00\4&6DA7DE7&0&0058
Adapter Type ATI Radeon Graphics Processor (0x95C5), ATI Technologies Inc. compatible
Adapter RAM 256.00 MB (268,435,456 bytes)
Installed Drivers atiumdag,atidxx32,atidxx64.dll,atiumdva,atiumd64.dll,atiumd6a.cap,atitmm64.dll
Driver Version 8.01.01.888
Driver c:\windows\system32\drivers\atikmdag.sys (8.1.1.888, 4.96 MB (5,203,968 bytes), 3/16/2009 3:34 PM)
Resolution 1680 x 1050 x 59 hertz

Processor:
Intel Core2 Quad CPU Q6700 @ 2.66GHz, 4 Cores, 4 Logical Processors

[jrweiss]

First, read through this thread for a lot of ATI, Catalyst 9.x, Core 1.24, and Environmental Variable information: viewtopic.php?f=51&t=9162

Second, since there is relatively little info available on the lower-end GPUs, you may have to experiment. In the past, 256 MB gfx cards did not fare well, so this case may be consistent.

Third, you cannot choose WUs, since GPU WUs are apparently only assigned on an ATI vs nVidia basis. The "big work units" and "advanced methods" switches have relatively little effect, compared with the CPU client.

[bruce]

Third, you cannot choose WUs, since GPU WUs are apparently only assigned on an ATI vs nVidia basis. The "big work units" and "advanced methods" switches have relatively little effect, compared with the CPU client.

It is true that you cannot choose WUs. It is true that ATI vs. nVidia WUs are currently assigned separately.

The "advanced methods" may influence the WU selection process, but it also may not. If there are no "advmethods" WUs being distributed, you'll get standard WUs. If there are some, you'll get a preference for them but there's still no guarantee.

Whether you get advmethods WUs or not will not change the behavior of your system in a predictable way. [Note: All standard WUs were once advmethods WUs,]

[toTOW]

The 3450 is a very low end chip ... you might have to run your machine 24H/24 to make deadlines on this GPU ... (and I don't know if it will be able to make deadlines even in this case)

[bebop]

I run the GPU core 24/7. Can someone elaborate on the advanced methods if it is anything more than the flag below (-advmethods)? I've not seen anything effective so far to help get a WU done in time.

I have the newest Catalyst Control center version, GPU version 6-23, and the latest core ATI 1.24

I'm using the following flag (some already attempted):
-forceasm -local -advmethods

See any issues with those, or one I should remove? I thought there was an option to limit things to 'small' 'large' or 'standard' - but I can only see it on the CPU version, not the GPU.

Also I changed the Windows Vista settings advanced environmental settings to:
BROOK_YIELD = 2
CAL_NO_FLUSH = 1

If anyone has any other ideas, I'm all ears. You'd think that since my GPU is supported, I could at least finish a WU under the time limit. Has anyone had any better performance overclocking the GPU in the ATI control center?

[jrweiss]

The FLUSH_INTERVAL setting may make a difference, too. Is your GPU running at 98-99% load, per the CCC?

[bebop]

FLUSH_INTERVAL = 32 (for now). I am running the GPU at ~97%, per the CCC. What next?

It looks like I'll be 5 hours overdue (still out 2 days and 6 hours) at this rate on the current WU, per the F@H Monitor utility. This is close than ever, but still not under the limit. I'm at about 29% complete with the current WU.

[bebop]

No help so far that has made any difference. Anyone have any ideas?

I got to 96% and then the final deadline was 1 minute away - so I set the clock on Windows back 7 hours to make sure I finished - instead of getting 0 points after 3 days. We'll see if it registers.

[jrweiss]

If you have the GPU load near 100%, there's little else you can do. Once the current WU finishes or aborts, let it try one more (hopefully a different project). If it still can't meet the deadline, you may have to give up on GPU folding with that card...

[bruce]

Setting the CPU clock will allow you to complete the WU but you won't get credit. The server uses it's own clock to enforce the deadlines. There's even an advanced setting in the configruation that will tell the client to ignore the local clock. That will allow you to work as long as you want on a WU that has already expired. The only reason to use the local clock is so that the client knows to discard a WU so you don't waste time working on something that has expired.

The deadlines are extremely important to the science.

Your hardware has to be fast enough to meet the deadlines. If it's not, then you've picked the wrong client or the wrong hardware. (The uniprocessor client has deadlines of a month or two and anything better than a 400 MHz Pentium can meet those schedules.)

[bebop]

My CPU is plenty fast 2.66 GHz (quad core). The GPU is overclocked to 720 MHz. It seems like a waste of time to run the GPU for 3 days and then miss the deadline by 2 hours.

Just not enough 'bandwidth' on the GPU processing power. If there was a good replacement video card I could get for $50 range, I might do that. I need one with HDMI for my home theater and one DVI for my computer. I'll start looking.

[Zagen30]

On Newegg right now there are a couple of 3650s going for around $50. I know those can make the deadlines, as the average PPD on them seems to be in the vicinity of 900 PPD (figure taken from legoman's chart on the ATI board). Also a couple of 9500 GTs, 8500 GTs (the 8500s at least seem to get around 700 PPD, 9500s might be similar) and one 8600 GT (1000+ PPD).

If you're willing to spend another $10 or so you could spring for a 3850 at around $60. I have one of those, and it gets around 1600 PPD (although since I upgraded a system that only had an AGP slot I had to pay twice as much And no, I don't have the money to build a new computer, much as I'd like to)

There are also some 4650s/4670s at the $60 mark. The 4670 seems to get around 1200 PPD, don't know about the 4650 but it's probably worse than the 4670.

[bebop]

I was thinking about this one for $80, now that I've looked around a bit:
http://www.newegg.com/Product/Product.a ... 6814102822

I wonder how many PPD I could expect from that?

[Zagen30]

From a 5 minute search of the forums I saw two reports of the 4830 getting somewhere in the vicinty of 1500 PPD. One of those people said it got between 1500 and 1800 depending on the project being run, which is completely normal for F@h.

Looking again I saw a report of 2000+, but that's road-runner, who probably has it overclocked significantly.

On the other hand, if you're willing to go Nvidia (don't know if you're loyal to ATI or something), the 9600 GSO can be found at around $80 and is supposed to get somewhere around 2700 PPD at stock settings. If you do go with the GSO, however, MAKE SURE YOU GET THE VERSION WITH 96 SHADERS, AKA STREAM PROCESSORs/SP's. Nvidia introduced a different version of the GSO that only has 48 SP's which folds much worse than the 96 SP version. One way to tell the difference is that the 96 SP model has either 384 or 768 MB of RAM while the 48 SP version has either 512 MB or 1 GB.

Currently TigerDirect has one for $80 before a $20 MIR. Of course, rebates are not all that reliable, but the base price of $80 is the lowest I've seen.

[toTOW]

I'd get a 4770 instead of a 4830 ... this is a brand new GPU on 40nm process, 640 SP, 750 MHz. It has the same number of SP as the 4830, but the higher clock yield to better performance, and the 40nm process helps with power consumption and heat and allows high overclocking. You can find them between $100 and $110 (and I think it will drop quickly as this board is supposed to replace the 4830).