I have all 12 threads crunching (6 phyisical + hyper-threading)
CPU: Intel 980X @ 3.33GHz

The extra time is due to the CPU overcommitment. As the majority of time on HD 79xx cards is when the task is run only on the CPU, when the CPU is overcommitted then the CPU time reported actually takes longer in wall clock time. This is because this CPU time is shared with the CPU tasks running. Basically a large proportion of the GPU task runtime is being spent waiting for CPU resources to be available.

This is the reason I previously said that allocation of sufficient CPU resources makes the greatest difference to the current beta HCC GPU task efficiency. Because it is the factor that makes the biggest difference in total task runtime.

From the Menu bar in Boinc Manager

Advanced > Preferences > Processor usage > On multiprocessor systems, use at most [91.67 ] % of the processors > OK

I think that will use 11 out of 12 threads and leave one free on a 980X CPU.

But there are no Beta GPU right now so maybe you should wait, unless you are running other GPU projects elsewhere.

Hello Platoon,
BOINC allows you to set up Device Profiles named Home, Work, or School. Also Default, but that might be overdoing it. Just select My Grid - Device Manager - Device Profiles - (Create a new BOINC Device Profile for *****). SAVE the Profile after editing "On multiprocessors, use: 50.0% of processors" or whatever you choose(?49.0%?), then assign it to your GPU systems.

We may be changing the website a bit to accommodate GPU computing. You might prefer to change the BOINC Manager on an individual GPU computer system by selecting BOINC Manager -Advanced - Preferences and editing the Processor Usage page to choose to ALWAYS run the GPU and the "On multiprocessors, use: XX% of processors". If so, remember to CLEAR that computer system if you want it to run the website preferences later instead of local preferences.

Remember that 50% on a hyperthreading CPU will only count physical cores, not virtual cores. 49% will leave a physical core without an assigned BOINC work unit.

Please tell me what I am not explaining clearly enough.
Lawrence

Based off Round 4: For what it's worth, here are a list of GPU times DO NOT BUY ANYTHING BASED OFF OF THIS, I WAS BORED AND WANTED TO SEE THE VARIOUS KERNEL TIMES, CPU TIMES, AND FINISH TIMES. YOU HAVE BEEN WARNED!!!


1) 295@1242=281.55 kernel, cpu=289.3, finished in 504 (local memory 16KB)

2) 450@1566= 389.23, cpu=493.36, finished in 497

3) 460@1024=219.045 kernel, cpu=291.11, finished in 290

4) 460@1440= 254.02, cpu=330.50, finished in 331

5) 460@1440= 241.99, cpu=321, finished in 321

6) 460@1600= 217.61, cpu=332.29, finished in 369

7) 480=1401= 98.94, cpu=157.10, finished in 159

8) 555M@1180= 631.29 kernel, Cpu=116 finished in 777

9) 560@1620=242.17 kernel, cpu=91.1, finished in 397

10) 560Ti@1645= 187.34, cpu=305.62, finished in 321

11) 560Ti@1645= 213.975, cpu=338.85, finished in 350

12) 560Ti@1800= 229.38, cpu=348.39, finished in 304

13) 570@1600= 121.43 kernel, cpu=202.08, finished in 202

14) 570@1700= 79.72 kernel, Cpu=153.95 finished in 154

15) 580@1800= 72.57, cpu=140.51, finished in 150

16) 590@1260= 108.79 kernel, cpu=172.38 finished in 173

17) 680@1110= 165.36 kernel, cpu=214.875 finished in 216

The 680 was mine, and all of them were returned as invalid, despite that they ran all the way thru. I switched drivers to the one other 680 which ran in hopes that this was the issue. The spread in kernel time for my 680 ranged from 165-235.