Late to the party but happy to help.
Late to the party but happy to help.
Originally Posted by MaxCFM
Thanks for helping!
Milestones Today
jeanguy2 passes 200,000
Hello folks,
Is the video card memory speed improtant for Einstein BRP5 WUs or is it like it was when we crunch ATI hcc1 WUs?
I had a rig running 2 X AMD crossfire 7970s at 1010 / 1375 which I have switched to 1010 / 900 without any apparant drawbacks.
Also what amount of GPU utilization are you guys getting? My GPU utilization with 2 WUs per card is in the 55 - 70% range. I am tempted to add one more WU to each video card but everytrhing I have read says stick to two Einstein WUs max per card? What do you think?
My 7870s run at 80% with 2 WUs. While most seem to say more than 2 shows no advantage it would not hurt to try. The 7970 is power house.
Everyone says you need 1 core per WU. I tried it in the past and saw no difference. Tried it again with different hardware, let 4 GPUs on 2 machines have 2 cores per GPU. After 3 days no difference to speak of. Back to .5 core per WU. WCG needs the cores and no benefit for E@H on my setups. Why my GPUs do not see the difference I do not know. I know the people saying 1 core per WU know what they are talking about.
From what I understand memory clocks are the least important on E@H. I read to OC everything else and when at max then up the memory. If you reduced it and saw no difference other things are probably what is holding your GPU back.
A lot depends on the chip set. On the projects #1 top host I'm running five BRP5 Wu's/7970 getting 98% usage. Motherboard is MSi big bang2. I've found the BRP4G app better on my older rigs.
Last edited by stoneageman; 09-29-2014 at 09:47 AM. Reason: Can't count. It's 5 tasks instead of 3
Thanks PG, I am also sticking to 0.5 CPU per Einstein GPU WUs for now. I have also noticed something that puzzles me. The 3930K CPU that I have runs at 4500 MHz and the 2600K CPU at 3800 MHz. When I compare the actual CPU time needed to complete a BRP5 WU, they both range between 1,300 to 1,450 secs. The GPU runtime for the 3930K is 10,600 to 10,800 secs while the GPU runtime for the 2600K is 18,200 tto 19,600 secs. Do Einstein @ Home WUs use the CPU througout the whole GPU WU processing rather than at the start and end like hcc1 WUs did?My 7870s run at 80% with 2 WUs. While most seem to say more than 2 shows no advantage it would not hurt to try. The 7970 is power house.
Everyone says you need 1 core per WU. I tried it in the past and saw no difference. Tried it again with different hardware, let 4 GPUs on 2 machines have 2 cores per GPU. After 3 days no difference to speak of. Back to .5 core per WU. WCG needs the cores and no benefit for E@H on my setups. Why my GPUs do not see the difference I do not know. I know the people saying 1 core per WU know what they are talking about.
From what I understand memory clocks are the least important on E@H. I read to OC everything else and when at max then up the memory. If you reduced it and saw no difference other things are probably what is holding your GPU back.
Thanks Sam, two of 7970 video cards that I have are MSI R7970 - 2PMD3GD5. I will let them run for another 12 hours with 2 Einstein WUs a piece to make sure that they remain error free with no invalids. I will then add one WU to each one to get their GPU utilization closer to 100%.
E@H work units use the CPU throughout the entire time.
I can only say that I saw significant performance difference when CPU threads were saturated while running BRP tasks. Thus freeing additional CPU core(s) has helped much.
You can make a simple test: watch the CPU usage of BRP tasks while running different amount of other CPU tasks - suspend them all, then resume one by one.
Once you get to a point where the CPU usage of the GPU task starts to drop, you know the CPU is saturated and that is going to hurt your GPU WU performance.
I'm using BoincTasks which shows this, you might also use other tools.
The point I was trying to make, and did not do a good job of doing, is to experiment a little as your results may vary. My systems seem to have a bottleneck. Not too worried as WCG is important to me. I not unhappy with what I am getting, even if it could be better.
Milestones Today
jeanguy2 passes 350,000
MaxCFM passes 30,000
Also, the team passes 450Million
Thanks a lot Mumak, I will use your above procedure to fine tuned my 2600K rig and find its Einstein GPU vs WCG CPU sweet spot. I am like PG and favor WCG work over Einstein work but I cannot let two 7970 GPUs crunch at only 50-60% of their potential.
Milestones Today
jeanguy2 passes 500,000
MaxCFM passes 50,000
Milestones Today
CeSinge passes 12,000,000
jeanguy2 passes 700,000
MaxCFM passes 60,000
Milestones Today
stoneageman passes 200,000,000
jeanguy2 passes 800,000
Milestones Today
Snow Crash passes 50,000,000
jeanguy2 passes 950,000
Yep, like this!
Nice
Milestones Today
jspace passes 20,000,000
jeanguy2 passes 1,200,000
Milestones Today
jeanguy2 passes 1,500,000
If I am looking at this right, from the Server Status page, 8 of us on this team are dominating a project with 37,317 active participants.
We need a few more crunchers!
Last edited by PoppaGeek; 10-05-2014 at 09:58 AM.
Posting Permissions
Reply With Quote
Bookmarks