Slicker
12-22-11, 10:59 AM
In the release notes of the current AMD APP SDK which is version 2.6, some of the key items mentioned are:
As of SDK 2.7, AMD CAL is no longer supported.
As of SDK 2.7, Windows XP is no longer supported.
When using Catalyst 11.7 drivers, some OpenCL applications use 100% of the CPU time. AMD aims to fix this issue in a future release of the Catalyst driver. Please upgrade your Catalyst driver as new releases become available. Unfortunately, 11.12 breaks Collatz at this time so using 11.6 or earlier is suggested.
As of APP SDK 2.6, AMD finally has async memory copies working in OpenCL although it requires setting an environment variable in order for it to work properly which is not good for BOINC apps. Async copies should speed things up a little. Collatz doesn't do that many copies (copy data to GPU, run GPU kernel thousands of times, copy data from GPU) compared to other apps like SETI (copy data to GPU, run single GPU kernel, copy data from GPU, repeat thousands of times). So, projects where the nVidia app is faster than the AMD app will probably get the most benefit from this - if they enable it and recompile the apps and require users to upgrade to Catylist 11.12 or higher.
As of SDK 2.7, AMD CAL is no longer supported.
As of SDK 2.7, Windows XP is no longer supported.
When using Catalyst 11.7 drivers, some OpenCL applications use 100% of the CPU time. AMD aims to fix this issue in a future release of the Catalyst driver. Please upgrade your Catalyst driver as new releases become available. Unfortunately, 11.12 breaks Collatz at this time so using 11.6 or earlier is suggested.
As of APP SDK 2.6, AMD finally has async memory copies working in OpenCL although it requires setting an environment variable in order for it to work properly which is not good for BOINC apps. Async copies should speed things up a little. Collatz doesn't do that many copies (copy data to GPU, run GPU kernel thousands of times, copy data from GPU) compared to other apps like SETI (copy data to GPU, run single GPU kernel, copy data from GPU, repeat thousands of times). So, projects where the nVidia app is faster than the AMD app will probably get the most benefit from this - if they enable it and recompile the apps and require users to upgrade to Catylist 11.12 or higher.