You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
GPUs have several characteristics that can be queried by the application. But there a lot of them, and most of them should be irrelevant to the user. Which ones are important and should be exposed to the user?
The following can help with the application logic:
I imagine we can expose them as methods/fields in the locale type. Currently, we can't do here in GPU kernels, so it may require us to do an interim solution first. (#22147)
I would add a variable identifying the device, e.g. something like "Tesla P100-PCIE-12GB" returned by nvidia-smi -- this would be very useful on a heterogeneous HPC cluster with multiple GPU types. Not sure if char cudaDeviceProp::name returns that information, but something along those lines.
GPUs have several characteristics that can be queried by the application. But there a lot of them, and most of them should be irrelevant to the user. Which ones are important and should be exposed to the user?
The following can help with the application logic:
multiProcessorCount
maxThreadsPerBlock
warpSize
(but I think we can do this without asking the device, see SupportwarpSize
or equivalent to query warp/wavefront size #23599)Maybe the following can help with diagnostics:
memoryBusWidth
/memoryClockRate
The text was updated successfully, but these errors were encountered: