• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Interesting info for those considering a GPU client machine

B

Bill Clo

Gawd
Joined
Jan 24, 2002
Messages
959
I did a little experiment to see how much power could be saved on machines that are candidates for conversion to the GPU client.

The machine tested is an Athlon 64 2.0ghz, 256mb RAM. It's a marginal folder on most Gromacs units, and only does a little better on Ambers, so I gave consideration to swapping it's motherboard out for one that's PCI-E compatible and adding an used X1900XTX card.

Baseline info:
Folding a P2124 unit, 2.0ghz. 113w used.
Underclocked to 1.2ghz 101w used.
Underclocked to 1.0ghz 97w used.

Normally underclocking it to 1.0ghz would result in much longer frametimes, on the order of twice as long. But since the GPU client isn't so picky as to the speed of the processor supplying it...

Undervolting the processor: left at 1.0ghz, 200FSB x 5 multiplier. Stock volts is: 1.45v.
1.4v 91watts used
1.35v 90 watts used
1.3v 88 watts used
1.25v 86 watts used
1.2v 84 watts used
1.15v 82 watts used
1.1v 81 watts used.

I didn't take it below 1.1v due to dimishing rate of return.

So for a dedicated GPU machine, I went from 97 watts to 81 watts by undervolting. Savings of 16w. Really I went from 113 to 81, with no real change in utility for the GPU client most likely.

It appears to me that those folk with machines capable of running the GPU client, faster than 1ghz, might realize some energy savings by slowing down their processor to 1ghz and underclocking it some. Why spend money on power if it isn't necessary?

32 watts saved equals $2.19/month, $26.28/year, $52.56 for 2 year Folding career. Offset by increase of 80w for the X1900XTX of course.

I'll update this when I get an appropriate motherboard and GPU for this processor.
 
You will find that the GPU client really does want at least 2.0ghz to fold fast. It will use more if availible but its a diminishing rate of return
 
Bill Clo said:
Have you any more info on this? I've seen info that suggests otherwise.
Click to expand...

With my GPU client by itself I got 7:45 per frame.

When I installed an SMP client running in Vmware using what's left of the cpu cycles I get about 9 minutes per frame. This is on an Oced to 3.03 C2D so with that said, I think 2 ghz should be the minimum.
 
I noticed a change in frame times on my GPU going from a X2 @ 2750 to a Opty170 @ 2850 and then to a X2 @ 2650. I'm a little too busy with work, Christmas and trying to maintain my recently rediscovered top 20 position in production to do any testing on it right now, but I'm sure you could figure out a pretty decent understanding of the scaling with some testing with a GPU running on the same CPU/System at 3 or 4 different clocks. It's probably not going to be like the old days folding a tinker core with a XP where it scaled directly in line with the clock increase. 20% higher clock would fold 20% faster.
One thing that I stated before is that a AGP card of the 1600Pro or 1650Pro variety may be a way to breath some life into some older folding boxen. XPs (that may very well have carried the team for a long time) and P4s just aren't what they used to be. ;) I was very tempted to buy a 1650Pro because they are around $100 just to try it out on a AGP/AthlonXP rig, but I sold the one I had running before I had a chance to do it.


 
I had previously done some very limited scaling testing. Ran a GPU client on a system with an X2 3800 OC'd to 2.4ghz. Re-ran it at 2.0ghz with less than 3 seconds frame time difference on the GPU client ( 6:06 at 2.4ghz, 6:09 at 2.0ghz). I haven't the time at this point to fool with that system and scale it down again. (it's super picky about settings, and it was a real cast iron bitch to OC in the first place; I don't WANT to mess with what's working finally :) ).

Anyone have more data to add to the pool? :)
 
Bill Clo said:
I had previously done some very limited scaling testing. Ran a GPU client on a system with an X2 3800 OC'd to 2.4ghz. Re-ran it at 2.0ghz with less than 3 seconds frame time difference on the GPU client ( 6:06 at 2.4ghz, 6:09 at 2.0ghz). I haven't the time at this point to fool with that system and scale it down again. (it's super picky about settings, and it was a real cast iron bitch to OC in the first place; I don't WANT to mess with what's working finally :) ).

Anyone have more data to add to the pool? :)
Click to expand...

I do have a question, this saved 32 watts you mentioned. Is that measured at the wall where you plug in or are you simply going by the math?
 
BillR said:
I do have a question, this saved 32 watts you mentioned. Is that measured at the wall where you plug in or are you simply going by the math?
Click to expand...

Measured with a Kill A Watt at the wall.
 
Update to the original post:

Got the X1900XTX running and tried a slower CPU speed.

Original config: A64 3200 @ 2.0ghz. Frame time w/GPU client: 6 min, 12 seconds. Using 207w power.

CPU reduced to 1.0ghz. Frame time w/GPU client: 6 minutes 32 seconds. 20 seconds/frame slower. Using 189w power.

Increase in time to complete a work unit = 33.33 minutes per w/u. Or basically an 4% increase in completion time for a power useage reduction of 9%. Saved power: 18w.
 
thanks for that update, but how many extra PPD could you get if you ran a CPU client in addition?
 
Final report on this system.

Here is some data, and I'll let those of you more skilled at data analysis have at it.

Using stock CPU voltage, frame time and power useage at various CPU speeds, GPU Client on an ATI X1900XTX (650core/775mem, ie, stock 3d speed). A64 Socket 754 1mb cache processor, 512mb memory.

2.0ghz 6:12/frame 207w power
1.8gz 6:16/frame 207w power
1.6ghz 6:18/frame 198w power
1.4ghz 6:22/frame 193w power
1.2ghz 6:26/frame 192w power
1.0ghz 6:32/frame 189w power

After graphing it all up, I found that the power useage curve dropped off sharply between 1.8ghz down to 1.4ghz, and then leveled off at 1.2ghz before dropping again at 1.0. I think 1.4ghz is the sweet spot here.

Data showing power consumtion versus increase in frame time:
1.8ghz 0% reduction power, +1.07% frame time
1.6ghz 4.3% reduction power, +1.6% frame time
1.4ghz 6.7% reduction power, +2.6% frame time
1.2ghz 7.2% reduction power, +3.76% frame time
1.0ghz 8.6% reduction power, +5.37% frame time

Next I tried undervolting the CPU. CPU speed selected is 1.4ghz. Stock CPU voltage is 1.45v. I ran Orthos with a simulated Gromacs work unit to avoid corrupting the GPU client work unit if there was a Early Unit End due to crashing. I know the data isn't strictly comparable, but it gives an idea of what kind of reduction to expect with undervolting.

CPU voltage power watts used
1.45 (stock) 118w
1.35v 108w
1.25v 103w
1.15v 99w
1.05v 95w

I left the CPU volts at 1.05v and started up the GPU client. No change in frame times, and it's been processing for an hour now with no noted instability. Total system power useage dropped to 166-168w (keeps fluctuating between the 2 readings). So I went from 207w to 168w (savings of 39w) with a modest reduction in production of 2.6% (766 Points per day to 747: 19ppd difference).

As a side note, before I did this round of testing, I purchased an Antec 430w high efficiency power supply and tested it with the system still set up for CPU Folding. Power useage dropped from 113w to 97w. This is from an older Antec power supply rated at about 70% efficiency to the claimed efficiency of around 83% with the new one). See here: http://www.silentpcreview.com/article684-page1.html

Hope this helps someone out and is informative.

 
You must log in or register to reply here.
Back
Top