The Scottish Boinc Team

So I have a Radeon R9 270X and I've configured Einstein to run mutiple tasks at the same time. Experimenting with the app_config.xml <gpu_usage> and <cpu_usage> options, setting gpu_usage between .2 and .5 to see how it affects the Activity reported under the Catalyst Control Center-Performance/AMD Overdrive tab. I have cpu_usage set to 1, have 4 cpu cores. Any lower and the Activity drops. I get the following activity levels based on gpu_usage.
gpu_usage .5 Activity avg 32% - 42%
gpu_usage .33 Activity avg 59% - 65%
gpu_usage .25 Activity 71% - 75%

Is the Activity reported on the AMD OverDrive tab accurate?
Do those gpu_usage results seem reasonable?
Is it worth trying to get better Activity, say 80-90%? If so, how?

Perhaps einstein is unique in requiring 3-4 tasks/gpu to get good gpu utilization. primegrid is averaging 90% with no special config. I had fun with the einstein challenge and learned a lot about configuring boinc apps, but not going to put any more effort into tweeking the einstein app. I'll let the remaining tasks finish, along with other projects tasks, but time to plan for the milkyway challenges coming up.

I used to spend daft amounts of time getting each project optimised. These days I am happy to let the automatics get on with it.

Einstein is unique. First off I would say, rather than focusing on gpu load, focus on time to complete.
Here are a few tips for Einstein
First Einstein allows you to set number of tasks per GPU in your account. Einstein@Home preferences.
GPU utilization factor of BRP apps and GPU utilization factor of FGRP apps. Changing the value from 1 to 0.5 runs 2 tasks. 0.33 - 3 tasks etc. Far easier than config files.
Secondly leave the cpu usage as is for most projects. Instead use the Boinc manager. Under Tools. Computing preferences. On multiprocessor systems, use at most....% of the processors.
The trick here is that it only sets for cpu work. Set this value to 50% effectively turns off hyper threading. Set to 75% and 3 cores will crunch cpu work, leaving 1 core free to run windows, the GPU etc. This will generally give you more through put as the cores will not fight each other for cache space. Again monitor how fast you finish units, not load on the cpu.
Once you understand this then you will see that the best way is to run 1 task / GPU. do 5 - 10 and note how long it takes to complete. Next do 2 tasks / gpu. again run a few and note times to complete. They will take longer, obviously, but if you can complete 2 tasks simultaneously faster than 2 tasks individually, then that is good. Do the same for 3 and see if that is better or worse.
BRP tasks and FGRP tasks are the only ones you should do this for. The other apps all need to run 1 at a time.

To optimise for Collatz see here http://www.dunadd.co.uk/seti/forum/view ... =26&t=1942

To optimise for Milkyway see here http://www.dunadd.co.uk/seti/forum/view ... 474#p16474

Another point to note is that Milkyway is double precision. Normally ATI cards rule here. I to have a Radeon R9 270X and was disapointed to find that the DP on it is artificially neutered by ATI to only 1/16 ( ie 1/16 of the number of stream cores )which makes it no better than my old 7750's and 7770's. My 7970 by comparison is 1/4 DP. The R9 280 / 280 x is 1/4 DP The R9 290 / 290 X is 1/8
Had I known that at the time I would have bought the 280x. The 290 has better SP flops but the 280 wins hands down for DP though even these new cards don't match the 7970's. More details here http://en.wikipedia.org/wiki/AMD_Radeon_Rx_200_Series