AMDave
02-08-2012, 10:32 PM
Separation updated to 1.00
I've updated all of the separation applications to 1.00. For changes people might care about,
- The old CAL version is gone; it's replaced with the OpenCL application. On AMD/ATI GPUs (older than 79xx) it is using some hackery to use the same IL kernel as before so it should be as fast. However this also means the Radeon 38xx cards aren't supported with new stuff. - Radeon 79xx stuff should work - The occasional validate errors from empty / truncated stderr should stop - AVX will be used if available on Linux and Windows (64 bit only for Windows) - I've increased the default GPU target frequency so GPU stuff should make things less laggy on average. You can now also configure this with the web preferences now so you don't need to use app_info stuff if you want to play with that. - Partial workaround with high CPU usage with recent Nvidia drivers.*
As usual post problems you run into here. * It should cut down on the CPU usage a bit while not sacrificing too much. I would recommend not using it unless you are very unhappy with the CPU usage on Nvidia. There are options to change the polling mode if you want to lower the CPU usage further while not slowing it down. (--gpu-wait-factor (default = 0.75) and --gpu-polling-mode (default = - 2) work similarly to how they did with the old CAL one, but slightly different). With the default of -2 it will use mode -1 unless it is an Nvidia driver newer than the one that introduced the high CPU issue, where it will use mode 0. Mode -1 uses the correct waiting method, mode 0 use the correct waiting method with an initial sleep based on time estimates, and modes > 0 are a polling period in milliseconds. The wait factor is a sort of correction of the time estimate used for the initial wait. The default is 0.75, to wait for 75% of the estimated time before trying to poll.
More... (http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=2764)
I've updated all of the separation applications to 1.00. For changes people might care about,
- The old CAL version is gone; it's replaced with the OpenCL application. On AMD/ATI GPUs (older than 79xx) it is using some hackery to use the same IL kernel as before so it should be as fast. However this also means the Radeon 38xx cards aren't supported with new stuff. - Radeon 79xx stuff should work - The occasional validate errors from empty / truncated stderr should stop - AVX will be used if available on Linux and Windows (64 bit only for Windows) - I've increased the default GPU target frequency so GPU stuff should make things less laggy on average. You can now also configure this with the web preferences now so you don't need to use app_info stuff if you want to play with that. - Partial workaround with high CPU usage with recent Nvidia drivers.*
As usual post problems you run into here. * It should cut down on the CPU usage a bit while not sacrificing too much. I would recommend not using it unless you are very unhappy with the CPU usage on Nvidia. There are options to change the polling mode if you want to lower the CPU usage further while not slowing it down. (--gpu-wait-factor (default = 0.75) and --gpu-polling-mode (default = - 2) work similarly to how they did with the old CAL one, but slightly different). With the default of -2 it will use mode -1 unless it is an Nvidia driver newer than the one that introduced the high CPU issue, where it will use mode 0. Mode -1 uses the correct waiting method, mode 0 use the correct waiting method with an initial sleep based on time estimates, and modes > 0 are a polling period in milliseconds. The wait factor is a sort of correction of the time estimate used for the initial wait. The default is 0.75, to wait for 75% of the estimated time before trying to poll.
More... (http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=2764)