Hello guys,

I want to know what would happen if a SP (Streaming Processor) do erroneously on a GPU? Can we design dispatcher/scheduler to avoid using these cores automatically? Or is it possible for software to handle this itself? Bear with me, I'm a newbie in GPU world!

Thanks.

Sure, I'll be glad to bear with you as a newbie, but ... I'm not sure I understand your question. Can you explain any further what you mean and what you are talking about?

Dan

Thank you very much for your reply, Dan.
Let me rip my question into two pieces :

1. What would happen if a core in a GPU does erroneously ?
Imaging there are some SPs in GPU which are faulty. I want to know what would happen if the dispatcher/scheduler assign them a job. I mean is it going to come up with a wrong result, or hardware/OS will infer that there is something wrong.

2. If there is already a hardware/software approach to avoid using these defective cores, please tell me about it.

I hope now my question make sense.
Thank you very much for your help.

While I have worked with others who have used GPU's, I have little experience with them myself. May I hazard a guess?

My first guess is that if there's a bug in the code sent to the GPU, it will act like most any other bug. Sometimes it will produce the wrong answer. In these cases, you might get strange ... objects, shapes, graphics, on your screen. If applied to a scientific problem, as opposed to the display, it might create noticable artifacts, or it might not. Indeed, this might have happened within a global climate model ... and how would we know? At least equally likely to both of the above would be a NULL-pointer type of error that would halt the program with an exception. If "properly" handled, that will halt your "program" with an error to be returned to the user.

Or ... are you referring to if the core has a bug within it? Hopefully, testing would find this bug before it was delivered. A good test-suite is very important for this purpose. That should capture the case of the entire CPU being bad.

But ... what if just one of the processor's within a GPU is bad? Imagine creating a GPU with NxN sub-processors within it, and allowing a percentage of them to be bad. If you could still deliver a working GPU, having routed around the bad subprocessors, you might manage to have a higher yield from the foundry.

Just ... some thoughts. The bottom line, though, is that I'm not sure I even know the answer to your question.

Dan

Thanks for your reply, Dan.

The last case you mentioned, was exactly what I'm talking about. What would happen if we design a GPU which can route processes around the bad and good subprocessors ?
I'm talking about a new architecture which software/driver can choose exactly between cores (i.e. Streaming Processors). So if some of our subprocessors got damaged during the years, or just after manufacturing we can still manage to use GPU without any problem. What do you think ... is it possible theoretically ?

I think I've even seen it done, once, but I never knew any of the details about how they did it.

Dan