While running WUs on an ATI 4870, it takes about 3 min to complete it. However, after activating the card build in on the motherboard (HD3300) times are longer by 50%. Two WU are done in parallel but they take about 9 minutes to complete. This strange behavior was not detected before but it was alternating WU from other project. Any ideas how to fix it?

The HD3300 cannot do MW units as it is not capable of double precision, it looks like your times are okay from what I can see though running right around 3 minutes or so.

There is not mention of the 3300 listed in the returned units.

---

While running WUs on an ATI 4870, it takes about 3 min to complete it. However, after activating the card build in on the motherboard (HD3300) times are longer by 50%. Two WU are done in parallel but they take about 9 minutes to complete. This strange behavior was not detected before but it was alternating WU from other project. Any ideas how to fix it?

What is likely happening is that BOINC is sending off the tasks to the 3300 but they are running on the 4870 so you are trying to run two tasks on the one GPU. If we did not have the Strict FIFO rule you might be able to get away with running MW tasks on the one GPU and Collatz on the other ... sadly UCB is still not listening to us and thus we have these situations where with asymmetrical GPUs (common and likely to get more common as time wends on) tasks have to be carefully scheduled on the correct hardware.

You might want to try the same experiment on Collatz and if it does work, petition on the BOINC Alpha mailing list for them to make the corrections ...

The HD3300 cannot do MW units as it is not capable of double precision, it looks like your times are okay from what I can see though running right around 3 minutes or so.

There is not mention of the 3300 listed in the returned units.

What is likely happening is that BOINC is sending off the tasks to the 3300 but they are running on the 4870 so you are trying to run two tasks on the one GPU. If we did not have the Strict FIFO rule you might be able to get away with running MW tasks on the one GPU and Collatz on the other ... sadly UCB is still not listening to us and thus we have these situations where with asymmetrical GPUs (common and likely to get more common as time wends on) tasks have to be carefully scheduled on the correct hardware.

You might want to try the same experiment on Collatz and if it does work, petition on the BOINC Alpha mailing list for them to make the corrections ...

Thank you for the help. I have test the issue further and it seems that MilkyWay's WU "synchronize" at the start while Collatz' WU do not free the Device after reaching 100%. The dual device setup works fine when one project use one Device while the other use the remaining one however.

On average the crunching is slowed down and I am receiving less credits so I am disabling the HD 3300.

The bug, if any, does not seems to belong to BOINC nor to the project's apps but to the underlying code (libraries or runtime) the apps use. This code seems to allow one WU to see another's WU Device and improperly allow apps running on different devices to synchronize.