Yes, I was going to notify you the changes introduced.

I can't figure out why there's an outlier in the ranking. Let's keep an eye on it to see whether the error is reproducible and traceable.

After I've crunched one short workunit on my GTX 980 Ti (just to get to the top of that list too ;) ) I've disappeared from the "Top average performers (last week long runs)" list. I think this is a sign of the presence of another bug...

Thanks for noticing. Should be fixed and reliable now.

Please do not hesitate to post any error you find. When we released the performance tab we were also busy with our scientific projects and therefore we didn't have time to do a proper testing.

Thanks for noticing. Should be fixed and reliable now.

Thank you!

Hi Gerard, it's the performance tab fetishist again.
My astonishingly fast GTX 980 Ti host did another impossible trick now:
It got to the top of two batch with the same workunit:
GERARD_CXCL12_CHLKDER_mol4: Result ID=14910783


GERARD_CXCL12_CHLKDER_mol42: Result ID=14910783


Next time I will show you that it's trained to jump thru a ring of fire :)

or it gets automatically updated every time you make a change in the computer?

Yes.
But for hosts with mixed GPUs only the best one is showed (and used by default) by the BOINC manager

Is this an accurate statement? I had 2 980 card in 2 different identical computers except for Windows version. (Identical in hardware.) When I put a 980TI in the one it read as 2 980s on the Volunteers-->Hosts page. When I swapped that card for the 980 in the other PC (because of the location and airflow around the first one and heat) the second one now reads 2 980TI cards. Again, the only difference is one is Win7Pro x64 and the other is Win8.1Pro x64.

I have noticed another oddity in BOINC and as a result on the Performance page. The 980TI card has 6GB of RAM onboard and it reports as 4GB in program and on that page. This is not a problem that I can tell unless BOINC could use the extra RAM and simply does not based on its reporting or whatever. I just like to see accurate reporting and information and that would be my fetishing. lol

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org

That is a weird artifact indeed. Possibily caused by the "parsing" engine I'm using. I am not able to isolate the problem right now because the database got updated, if it happens again please let me know.

1) Multiple GPUs

This is a well-known oversimplification by BOINC. The BOINC client used by volunteers identifies GPUs individually. By default it only uses the 'best' GPU of each vendor (plus close matches): in a computer with mixed GPUs from the same vendor, the user can (and most activists do) direct BOINC to 'Use all GPUs'.

The over-simplification occurs on the BOINC server. Rather than listing each GPU separately, the server database summarises all the GPUs from each vendor into a single record - it displays the identification of the 'best' GPU, and a count of all GPUs, but discards the fine detail of multiple GPUs. This is an acknowledged weakness of the BOINC infrastructure, but there is no timetable yet for a re-write.

2) 4 GB memory displayed

Much of the CUDA support code is 32-bit only, and can't process numbers larger than 4GB.

I think 'best' GPU is based on the GPU models Device ID [hex] or BIOS version. GFlops was argued as a better option, but the opportunity to change these things has gone. The OS doesn't always get it right either, especially if you start swapping cards around, without un-installing, wiping and reinstalling the drivers. You can even end up with different drivers for similar cards.

The acemd app reads the details of the card directly - independently of Boinc, and the OS. Not sure if any 32-bit code would impact on acemd or not, but I haven't seen a task use over 1.5GB GDDR.

---

FAQ's

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

I think 'best' GPU is based on the GPU models Device ID [hex] or BIOS version. GFlops was argued as a better option, but the opportunity to change these things has gone. The OS doesn't always get it right either, especially if you start swapping cards around, without un-installing, wiping and reinstalling the drivers. You can even end up with different drivers for similar cards.

From client source code (gpu_nvidia.cpp):

// return 1/-1/0 if device 1 is more/less/same capable than device 2.
// factors (decreasing priority):
// - compute capability
// - software version
// - available memory
// - speed
//
// If "loose", ignore FLOPS and tolerate small memory diff

So the 'best' GPU (in the NV case, which concerns us here) is the one with the highest compute capability, and the others listed are tiebreakers.

And the 2nd runner up is....

exactly the first runner up.

I'm getting excited now.

Since I have not been cleared from the 3rd place when the numbers have changed, it could mean that I can occupy all 10 places simultaneously someday.

...but it won't happen very soon.

OT

I think 'best' GPU is based on the GPU models Device ID [hex] or BIOS version. GFlops was argued as a better option, but the opportunity to change these things has gone. The OS doesn't always get it right either, especially if you start swapping cards around, without un-installing, wiping and reinstalling the drivers. You can even end up with different drivers for similar cards.

From client source code (gpu_nvidia.cpp):

// return 1/-1/0 if device 1 is more/less/same capable than device 2.
// factors (decreasing priority):
// - compute capability
// - software version
// - available memory
// - speed
//
// If "loose", ignore FLOPS and tolerate small memory diff

So the 'best' GPU (in the NV case, which concerns us here) is the one with the highest compute capability, and the others listed are tiebreakers.

Thanks for the recap. Was only thinking about the second step, and thought the software version was the GPU's Bios version which is recorded in hex. Assuming it is the GPU Bios version, that would explain why we sometimes see smaller cards listed instead of the biggest and most capable cards. When smaller (GFlops) cards have the same compute capability but are released later with a more recent Bios they would be seen as the most capable card by Boinc.
You can also have a small Card with low GFlops reported as the best card when the smaller card is from a newer generation, but isn't more powerful. For example, a CC5.0 (GM107) GTX750 would be reported ahead of a big CC3.5 (GK110) such as a GTX780Ti [which is 3 times as powerful as the GTX750].
The amount of memory isn't a great way of identifying the 'most capable' GPU either. A lesser GPU (GTX750) could have 2GB or 4GB GDDR while a more powerful GPU (GTX750Ti) could have 1 or 2GB.
Speed doesn't say much also. Bigger cards typically run at lower frequencies, but have more shaders.
Not saying GFlops would be preferable as you would have to segregate into at least two categories of GFlops, FP32 (SP) and FP64 (DP), as different projects use or don't use DP. Then there is half precision which might change things further.

In reality the whole purpose and method of identifying GPU capability might need to be revisited. Why is Boinc doing what it is re GPU's? It's mostly scheduling and reporting. Would identifying cards as being ATI/AMD, NVidia or Intel be sufficient given that most apps are better equipped to identify GPU's?

Driver versions are probably more important than speed - most projects apps require a specific driver or range of drivers to work.
GDDR amount is important too as most projects require a GPU with sufficient RAM to run certain apps or tasks.
Bios versions might be helpful in identifying rogue cards but was probably more important back in the day of the GF200 - GT200-103-A2 vs GT200-105-B3 for example could identify 65nm from 55nm.

---

FAQ's

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

For some reason I'm not being ranked at all on the performance tab. This WU I did in 10.93 hours according to the performance tab, which according to the rankings, should be good enough for 25th place, yet 25th place on the WU chart itself is listed as belonging to (Ryle) with 11.02 hours. I completed said WU almost a week ago, so there should have been plenty of time to update, yet I'm not listed

I think it's because you've completed this WU a little over its deadline, and it has been assigned to a different host.

For some reason I'm not being ranked at all on the performance tab. This WU I did in 10.93 hours according to the performance tab, which according to the rankings, should be good enough for 25th place, yet 25th place on the WU chart itself is listed as belonging to (Ryle) with 11.02 hours. I completed said WU almost a week ago, so there should have been plenty of time to update, yet I'm not listed

The same thing happened to me a few weeks ago. It was a GIANNI WU. It finished within 24 hours. Though, if I remember it correctly, a previous host finished with an error on it. I didn't think it was noteworthy back then, so I didn't post at the time. Apparently, the system has another bug.

[quote]For some reason I'm not being ranked at all on the performance tab. This WU I did in 10.93 hours according to the performance tab, which according to the rankings, should be good enough for 25th place, yet 25th place on the WU chart itself is listed as belonging to (Ryle) with 11.02 hours. I completed said WU almost a week ago, so there should have been plenty of time to update, yet I'm not listed

The performance tab doesn't always update immediately you return a WU

[quote]For some reason I'm not being ranked at all on the performance tab. This WU I did in 10.93 hours according to the performance tab, which according to the rankings, should be good enough for 25th place, yet 25th place on the WU chart itself is listed as belonging to (Ryle) with 11.02 hours. I completed said WU almost a week ago, so there should have been plenty of time to update, yet I'm not listed

The performance tab doesn't always update immediately you return a WU

I did say that it had been a week.