Device HD7970.
Catalyst 12.4:
CL_DEVICE_MAX_MEM_ALLOC_SIZE: 524288KB
CL_DEVICE_GLOBAL_MEM_SIZE: 2048MB
Catalyst 11.12:
CL_DEVICE_MAX_MEM_ALLOC_SIZE: 786432KB
CL_DEVICE_GLOBAL_MEM_SIZE: 3072MB
Why was memory limit reduced? Is it regression or some kind of firepro marketing?
from release notes. so it is intentional.
In Windows and on Linux or Southern-Islands-based platforms, the OpenCL runtime reports the total amount of physical memory for boards with up to 2 GB.
But what's the reason behind this restriction?
I was able to use all 3Gb local memory with catalyst 11.12 and GPU_MAX_ALLOC_PERCENT. Now it's broken. Buffers large than 1.5Gb allocated in system memory. Is there any way to force local gpu memory allocation?
Same question here: I have two stock 7970's sharing via MPI, with a third card for the Win7 GUI. With Catalyst 12.8 and Sdk v2.7, clGetDeviceInfo(... CL_DEVICE_GLOBAL_MEM_SIZE, ...) reports only 2GB. Setting GPU_MAX_HEAP_SIZE to 100 does not seem to make a difference (I've not yet tested beyond 2GB).
Assuming the new FirePro cards (with 6GB GDDR5 ) with not have their memory disabled, will a future driver enable all memory on GCN cards, or will mere Radeons be identified, and knobbled for being cheap 😐
Actually all ~3GB could be used without problem. clGetDeviceInfo returns "safe" size.
You just can not allocate one buffer larger than 500Mb. But you could allocate five buffers, 500Mb each.
Thanks for the optimism! I'll get a chance to test in a few days, and see what breaks
Did some investigation and found this http://devgurus.amd.com/thread/159645
Maybe it will be usefull............