Folding Forum

This week I upgraded my GPU. The old GPU, an HD5550, had about a day left to complete its WU, and I estimated that I could swap in the new one and, even if there was a problem and I had to roll back, there would still be enough time for the old GPU to complete the WU.

This didn't happen though. When the new card was installed, FAH client just discarded the old WU and got a new one.

Is there anyway to avoid this, other than simply waiting for the current WU to complete?

Also, is this really desirable default behaviour from the client? Surely the new GPU should just finish the current WU?

I've had the same problem. Rarely I have been able to avoid the loss of the WU but I don't upgrade GPUs often enough to remember what works and what doesn't. I always do a FINISH before changing GPUs, avoiding the issue.

Also, it's a good idea to reinstall drivers and reinstall FAH since sometimes the new GPU is properly recognized and often it is not. (FAH simply assumes that the hardware will not change. I don't think that's a common enough situation to spend a lot of Development time enhancing the client and debugging the result.)

To answer your last question first - in general, no it should not. There is no certainty that a different piece of GPU hardware will process the WU the same as the prior one, especially if different drivers are involved. There are other factors as well that could cause the WU to be corrupted, possibly not in a detectable way.

So best practices is to use the Finish command to complete any current work before changing hardware configurations. In the case of changing GPU's, a further step is to reinstall the client after installing the GPU and its drivers.

Ok fair enough. I suppose if the WU is compiled for a specific core then it would be unpredictable on another.

The ATI HD5550 used ATI's Terascale 2 architecture, more modern AMD cards use GCN, so the drivers should be quite different.
https://en.wikipedia.org/wiki/TeraScale ... eraScale_2
https://en.wikipedia.org/wiki/Graphics_Core_Next