cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

peccable
Journeyman III

Firepro W7000 and OpenCL performance

I've got a workstation (Xeon E5-1650 v2, 64GB RAM, Firepro W7000) and use it for OpenCL coding. Now, according to the specs, the GPU should be way faster than the CPU for highly parallelizable tasks. However, the CPU performs about 20-30 percent better than the GPU in cases where there are on the order of 1k double precision floating point operations in a kernel (of course care is taken to ensure that memory access is sequential and that work items does not try to access the same memory all at once).

I've also noted that OpenCL reports less available memory for the GPU than the 4GB it is supposed to have. Here is the log from the OpenCL initialization:

Platform initialization:

      num platform ids: 1

      platform version: OpenCL 1.2 AMD-APP (1411.4)

      platform profile: FULL_PROFILE

         platform name: AMD Accelerated Parallel Processing

       platform vendor: Advanced Micro Devices, Inc.

     available devices: 2

Device 0; Firepro W7000:

  device address space: 32

   max work item sizes: 256, 256, 256

        device profile: FULL_PROFILE

OpenCL driver version: 1411.4 (VM)

        device version: OpenCL 1.2 AMD-APP (1411.4)

    device memory size: 3072 MiB

  device compute units: 20

   max work group size: 256

Devince 1; Xeon E5-1650:

  device address space: 64

   max work item sizes: 1024, 1024, 1024

        device profile: FULL_PROFILE

OpenCL driver version: 1411.4 (sse2,avx)

        device version: OpenCL 1.2 AMD-APP (1411.4)

    device memory size: 65459 MiB

  device compute units: 12

   max work group size: 1024

So does anyone know what the issue might be here?

0 Likes
1 Reply
chm
Staff
Staff

Hi,

you can use set the environment variable GPU_FORCE_64BIT_PTR to 1 to enable 64 bit pointers and to see the entire available memory. See:

http://devgurus.amd.com/message/1286769?tstart=0

Chris