Request for Clarification about Supported AMD GPUs
Posted: Tue Apr 14, 2020 12:57 pm
I have read a couple of places here that supported amd GPUs need double precision support, as noted by the columns in the wikipedia list of amd gpus.
Because AMDs branding and hardware IDs can be confusing to everyone*, I was wondering if there was a more specific indicator?
It seems like the a compatible card needs to support a specific double precision OpenCL extension for the GPU in question: cl_khr_fp64.
The presence of this extension can be checked a couple of ways: the output of clinfo in linux or windows, or using fahbench to execute a double precision benchmark.
Another line in clinfo that might be just as relevant is:
I think if it's 0/0 it's not supported.
Is either of these an accurate descriptor? I guess it makes it easier for driver/card testing but harder for folks shopping for used cards because the driver/api feature documents aren't clear.
*(I spent a while thinking that an radeon hd 8670d was GCN 1, but it's terrascale! The 8670 and 8670M are GCN though. *facepalm*)
Because AMDs branding and hardware IDs can be confusing to everyone*, I was wondering if there was a more specific indicator?
It seems like the a compatible card needs to support a specific double precision OpenCL extension for the GPU in question: cl_khr_fp64.
The presence of this extension can be checked a couple of ways: the output of clinfo in linux or windows, or using fahbench to execute a double precision benchmark.
Another line in clinfo that might be just as relevant is:
Code: Select all
Preferred / native vector sizes
...
double 1 / 1
Is either of these an accurate descriptor? I guess it makes it easier for driver/card testing but harder for folks shopping for used cards because the driver/api feature documents aren't clear.
*(I spent a while thinking that an radeon hd 8670d was GCN 1, but it's terrascale! The 8670 and 8670M are GCN though. *facepalm*)