Just curious as to how the credits are applied to WU's
Previous Santi's than this one (see below)..gave 20550.00 per WU but the one's listed below give 18300.00 per. Approx. same amount of CPU and GPU time.
Just wondering.

Regard's,

John

http://www.gpugrid.net/workunit.php?wuid=5143983

Yes, there appears to be a credit leak!

The SANTI_MAR WU's pay around 16% less than the SANTI_BAX WU's did.
The NOELIA_DIPEPT WU's are even more stingy (about 31% shy).

I409-SANTI_bax2-21-32-RND7270_2 5134632 4 Feb 2014 | 7:57:01 UTC 4 Feb 2014 | 17:48:46 UTC Completed and validated 32,501.28 32,155.69 154,050.00 Long runs (8-12 hours on fastest card) v8.15 (cuda55) [409K/day]

470x-SANTI_MAR422cap310-0-32-RND8919_0 5141239 6 Feb 2014 | 10:32:17 UTC 6 Feb 2014 | 19:16:55 UTC Completed and validated 28,378.57 11,605.68 115,650.00 Long runs (8-12 hours on fastest card) v8.15 (cuda55) [352K/day]

tyrglux6x44-NOELIA_DIPEPT1-1-2-RND3665_0 5143384 6 Feb 2014 | 3:16:34 UTC 6 Feb 2014 | 11:22:11 UTC Completed and validated 25,796.40 13,062.85 93,000.00 Long runs (8-12 hours on fastest card) v8.15 (cuda55) [311K/day]

I think some of the factors considered are task length, gpu usage and possibly CPU usage?!?

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

If the factors are as mentioned ..Yes there is a credit leak..
The Santi_Mar WU's appear to take about as much GPU and CPU time As the Santi_Bax WU's, but give less credit's.
That is why I was asking what determined credit per WU.
At least I am not the only one who has noticed.
Thanks for the reply.

Credit / WU is calculated based on runtime of a WU on one of our machines. So we run a WU on one of our GPUs check time required to run and multiply by credit/time. Then there is also the bonus for the long tasks and also the fact that we usually round the credits up. Additionally we sometimes submit the WU's with the wrong credits out of a small mistake in a input file (but we usually catch those).
So there can be many factors (different GPUs/our rounding/our mistake) that cause a discrepancy in the WU credits but we try to keep it as low as possible. Considering that users cannot pick the WU's they want to crunch I assume the credits average out.

Thanks for the reply/Info.
Let's hope the assumptions are correct

Hi!

And what about the bonus time?
I'm not sure, but if we send back a WU in les than 24 hours, the credits are bigger.
I'm right?

Yes 150% for WU's returned inside 24h, 125% for WU's returned between 24h and 48h, 100% thereafter (unless you get a resend and there is an issue - you get half credits if the original WU is returned, before you can complete the resend).

WU's are resent soon after the receipt of a failed run, or 5days after the WU is initially sent, if not already returned (this use to be 3).

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Yes indeed. Here is the FAQ: http://www.gpugrid.net/forum_thread.php?id=2572

OK, thank you for the link.

But, it's a pitty, I need 25h to give back a Noelia with my NVIDIA GeForce GTX 650 Ti...

But, it's a pitty, I need 25h to give back a Noelia with my NVIDIA GeForce GTX 650 Ti...

This looks little slow. I have two different GTX460 with 1Gb memory running under XP with the 334.89 drivers that take only around 16-17 hours on the last Noelia WU's they ran.

Card 1
http://www.gpugrid.net/result.php?resultid=7837842
http://www.gpugrid.net/result.php?resultid=7812020

Card 2
http://www.gpugrid.net/result.php?resultid=7813942
http://www.gpugrid.net/result.php?resultid=7811015

If setup dedicated 24/7 and crunch even when computer is in use, I would think the 650Ti could do <24h turn around.

[AF>Libristes] cottesloe, Your last Noelia WU took 81,042.66sec to run. This is less than 24h (86400sec). If you reduce your cache (Runtime Buffer) to 0.01 that should allow you to complete Noelia WU's inside 24h. The other WU's take longer 90K sec. If you reduce the number of CPU tasks you run it might decrease GPUGrid WU runtime, but that's just a guess. Other than the above, or upgrading it, all I can suggest is to increase fan speed, which would only prevent the card from downclocking due to heat (probably not an issue), or try to overclock the card (your on your own there).

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

[AF>Libristes] cottesloe, Your last Noelia WU took 81,042.66sec to run. This is less than 24h (86400sec). If you reduce your cache (Runtime Buffer) to 0.01 that should allow you to complete Noelia WU's inside 24h. The other WU's take longer 90K sec. If you reduce the number of CPU tasks you run it might decrease GPUGrid WU runtime, but that's just a guess. Other than the above, or upgrading it, all I can suggest is to increase fan speed, which would only prevent the card from downclocking due to heat (probably not an issue), or try to overclock the card (your on your own there).

Remember that you have to allow extra time for downloading and uploading, over and above the recorded elapsed (computing) time, if the entire "issue and return" cycle is to fit within the 24 hours.

If you reduce the number of CPU tasks you run it might decrease GPUGrid WU runtime, but that's just a guess.

It should be a good guess..
I'll try this next week...
Bye

I let one CPU free and I can see the diference...
Around 17h to crunch a Noelia, vs 25h before...

No comment...

Bye and good wishes from France!

It was 120,000 credit for long runs about a week ago and now rarely see 120,000 credits. Every time I get my rig going well, things change for the worse. 50,000 by 50,000 will take a while to climb the ladder.

It was 120,000 credit for long runs about a week ago and now rarely see 120,000 credits. Every time I get my rig going well, things change for the worse. 50,000 by 50,000 will take a while to climb the ladder.

You should take account of the running time of the given workunit. Shorter workunits generate less credit, but they do it more often.
However there is a small (~10%) variation in the credit/time ratio of the workunits.

I would prefer a metric in which the credits per unit of computing were comparable to other BOINC projects that don't use GPU computing.

---

I would prefer a metric in which the credits per unit of computing were comparable to other BOINC projects that don't use GPU computing.

We all would.
But it's much harder to measure (or to calculate) the actual flops done by a GPU task than a CPU task, because the GPU does many calculations simultaneously. As there is no standard for calculating the actual flops done by a GPU task (and probably there couldn't be such at all), different projects give different amount of credit per GPU time. It makes hard to put different GPU projects in comparison (especially with CPU projects), so the cruncher should not judge the amount of scientific contribution only by the credits given for GPU work. It's like comparing apples to bananas. We are crunching to aid research, not for the credits, right? After all, you couldn't do anything with the credits earned at any project, it's just for comparison between the contribution of the users of a given project, but not for the comparison between different projects. However, the much lager amount of credits earned by a GPU task correctly reflects the computing capability ratio between GPUs (parallel computing) and CPUs.

Measuring CPU performance is fundamentally flawed. The actual usefulness/performance is related to both the CPU and the task. Nowadays some instruction sets offer significant improvements for some work type over older instruction types, but others don't. Then there is the supporting hardware (RAM/disk/bus/chipset). Faster RAM significantly improves performances for some WU's... The GPU is different again, and trying to compare NVidia and AMD is a waste of time; they are designed to do different things.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I would prefer a metric in which the credits per unit of computing were comparable to other BOINC projects that don't use GPU computing.

Be careful what you wish for, 'creditnew' could come to the projects using gpu's too if Dr. A has his way. That would most likely mean a LARGE drop in credits for most workunits at most projects. 'Creditnew' is Dr. A's idea of synchronizing credits across all projects, so equivalent work gives equivalent credit at EVERY project. That would mean no more 133,950.00 credits units!!