Ok, let's assume that you are getting better number-of-results per time on the aggregate compared to a single GPU-WU done one at a time. How do we account for that better yield? There is no such thing as a free lunch: I strongly suspect that your power input must be higher, and that is inconsistent with your claim of '99% GPU utilisation and no significant fan noise nor power consumption increase'.
;
; andzgridPost#743
;

Unless things have changed or I have misinterpreted the information, maximum elapsed time is more like 10 hours
LINK https://secure.worldcommunitygrid.org/forums/wcg/viewpostinthread?post=401797

If I would do this for money, I would definitely make thorough analyses of my expenses as well as my production. Since I am not, I am satisfy with rough analysis based on small-sample observations and some logic

1. Noise is really subjective category and what I perceive as "no increase", for somebody else could be significant. So I am gonna to agree with you, that this part of my claim might be misleading. Anyway, I am repeating, that according to me, there is not significant fan noise increase
2. When I am running 32 concurrent GPU tasks, CPU utilization is way south of 100% (thus less power consumption). If I run only one GPU task + 7 CPU tasks, the CPU utilization on those 7 cores is 99-100% (thus higher power consumption). That is why I assume (and claim that) the system power consumption is not increasing significantly (if at all) if I run 32x GPU WU's compared to 1x GPU + 7x CPU.

Cheers and NI!

Maybe it is because of app_config IDK. But all "24 concurrent crunching" tasks finished in less than evil 1.5 hours, thus w/o errors

Hi,

Not sure you fully understand the use of these 2:

<gpu_usage>.5</gpu_usage>
<cpu_usage>1.25</cpu_usage>

You've set the the GPU to run 2 at the time and CPU allocation *per* GPU job to 1.25. The highest any WCG project uses as being single threaded, is 1 CPU processor thread, so reduce that value to 1.00 at least. But, given the 5650 CPU's you can try <cpu_usage>0.5</cpu_usage> and see if throughput per hour increases, meaning if one task is in GPU phase, the other task can use the CPU, in alternation. Then later try <gpu_usage>.25</gpu_usage> meaning 4 GPU tasks, using combined 4 * .5 = 2 CPU cores together. Increment that in small steps to find the biggest hourly production, without them going error/invalid.

edit: To add, with 2 such cards and CPUs, [16 processor threads], and <max_concurrent> of 24 you could theoretically pump the <gpu_usage>.5</gpu_usage> way down to .83333. Any spare instance what is not being run on the GPU will be running as CPU tasks, provided you've selected to also run HCC-CPU only tasks. The idea though is to experiment with the control values to find the maximum throughput maintaining valid results [Pending Validation is usually a good indicator that the task is OK]

edit2: I've never read [can't remember seeing it] if having 2 cards of same or different capability and using .083333 is a control value *per* GPU card or if it works for the combination i.e. 12 for the 2 cards together. With that value a total max of 12 GPU tasks of hcc1 would then mean 6 concurrent on each. anyone in the know from hands on?

edit3: made a oops. The .83333 value of course to read .083333, equivalent to 12 concurrent.