cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

Raistmer
Adept II

Unacceptable performance degradation after Catalyst update

Updated from Cat 11.2 to Cat 11.6

1) I got BSoD on Cat 11.2 uninstallation
2) After removal of Cat 11.2 and installation of Cat 11.6 drivers I see huge performance degradation in my application.

Moreover, OpenCL query now reports incorrect values for my GPU (HD6950)
Here what it was reported before and after update:


Number of OpenCL platforms: 1


OpenCL Platform Name: ATI Stream
Number of devices: 1
Max compute units: 22
Max work group size: 256
Max clock frequency: 840Mhz
Max memory allocation: 536870912
Cache type: None
Cache line size: 0
Cache size: 0
Global memory size: 2147483648
Constant buffer size: 65536
Max number of constant args: 8
Local memory type: Scratchpad
Local memory size: 32768
Queue properties:
Out-of-Order: No
Name: Cayman
Vendor: Advanced Micro Devices, Inc.
Driver version: CAL 1.4.1016
Version: OpenCL 1.1 ATI-Stream-v2.3 (451)
Extensions: cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_printf cl_amd_media_ops cl_amd_popcnt cl_khr_d3d10_sharing


after:

Number of OpenCL platforms: 1


OpenCL Platform Name: AMD Accelerated Parallel Processing
Number of devices: 1
Max compute units: 22
Max work group size: 256
Max clock frequency: 800Mhz
Max memory allocation: 209715200
Cache type: None
Cache line size: 0
Cache size: 0
Global memory size: 838860800
Constant buffer size: 65536
Max number of constant args: 8
Local memory type: Scratchpad
Local memory size: 32768
Queue properties:
Out-of-Order: No
Name: Cayman
Vendor: Advanced Micro Devices, Inc.
Driver version: CAL 1.4.1417 (VM)
Version: OpenCL 1.1 AMD-APP-SDK-v2.4 (650.9)
Extensions: cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_gl_sharing cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_popcnt cl_khr_d3d10_sharing


Wrong values are highlighted
(not sure about 840MHz because quite possible that card operated on 800MHz before too, but memory amount and single allocation block are definitely bad)

0 Likes
8 Replies
Raistmer
Adept II

More precisely, I see some improvement in CPU time consumption but huge performance drop in total elapsed time.

See picture:
http://img692.imageshack.us/im...2/apgraphhd6950.th.png

0 Likes
Raistmer
Adept II

Another update:
Sometimes I see timings almost the same as for Catalyst 11.2, i.e, same CPU usage and same elapsed time for application. But time to time it shows reduced CPU usage and vastly increased elapsed time with Catalyst 11.6 versus Catalyst 11.2.

What could be reason for such erratic behavior under 11.6 drivers and what can be done to avoid it (besides to downgrade to 11.2 again). With so many driver updates I though AMD was able to support HD6xxx GPUs in OpenCL adequately....
Maybe something should be changed in app itself to make it "Catalyst 11.6-compatible" ? Curent behavior just unacceptable...
0 Likes

Hi raistmer,

Can you please tell your system information.

What application are you using for duducing your results. Is it reproducible using some SDK samples?

0 Likes
Raistmer
Adept II

I use SETI AstroPulse application. Don't know if SDK samples will behave the same bacause of very short execution time of samples. My app runs ~hour.

System information is Q9450 as CPU and HD6950 with Catalyst 11.6 as GPU. OS is Vista x86.
On Catalyst 11.2 (as one can see from posted graph) execution times were very stable. Nothing even near to such deviation as for Catalyst 11.6.

0 Likes
Raistmer
Adept II

Any updates on this problem? I reinstalled drivers few times, even used Driver Sweeper to remove catalyst completely before reinstall - all this didn't help. In few times execution times were slightly lower than with cat 11.2, but in all other cases they were much bigger. That is, overall performance decreased significally with Cat 11.6 drivers vs cat 11.2 ones. Will try 11.7 beta now...

[btw, I see this behavior on both my applications, not just on one]
0 Likes

Under 11.7 situation became even worse. App started to throw exceptions due to too long running time...
0 Likes

Raistmer,

Can you please point to the app you are running. And please mention steps to reproduce the issue.

0 Likes
Raistmer
Adept II

Well, I think first step to reproduce it just to run CLinfo.exe provided with SDK.
So you see great decrease in max single buffer size and max available GPU memory size when going from 11.2 to 11.5, 6, 6 ? I see. (look initial posts).
0 Likes