7970 @ 1135/1890, LG 2, TC 22392: GPU 0: 72.0C 3413RPM | 714.6K/715.7Kh/s | A:0 R:1 HW:0 U:0.00/m I:20
Uhm. With LG 2 and TC 22392 I get: [2013-03-15 16:04:33] Maximum buffer memory device 0 supports says 805306368
[2013-03-15 16:04:33] Your scrypt settings come to 1467482112
[2013-03-15 16:04:33] Error -61: clCreateBuffer (padbuffer8), decrease CT or increase LG Ok, the above error seems going away using: export GPU_MAX_ALLOC_PERCENT=100
export GPU_USE_SYNC_OBJECTS=1 (I never ever needed this thing before). As always YMMV, just because it works on mine doesn't mean it will work on yours - motherboard, CPU, and ram actually matter with scrypt. However -g 1 is now almost mandatory with these higher TCs. I'm making 1 GPU thread default in scrypt in the next version (just made it into git). |
|
|
|
Found a nice awkward bug and the next version of cgminer will fix the high thread concurrencies limitation on 79x0 devices and the intensity 13 limitation.
Oh yeah, I did not check the name sorry. It makes a difference in reaper, not cgminer, I forgot to mention. But I am telling you the truth trust me  Also, very good work on cgminer, and if the next release will indeed lift the intensity/concurrency issue, i'll get back right away from reaper on one of my cards, because the stratum support absence is a pain, applying the proxy all the time. Keep up the cool work! I'm not sure you want to hear this, but the scrypt kernel in reaper is the same as in cgminer, and it also does not do anything with vector parameters for litecoin mining (strictly speaking, both kernels are fixed to 4 vectors internally). Anyway thanks. Test builds can be found on the cgminer forum thread. |
|
|
|
|
Sold 3 out of 4 of my 7970s already. It's an awesome time for 7970 resale.
|
|
|
|
For all you crazy scrypt/LTC mining fanatics, I have finally found the reason you cannot set very high thread concurrencies or intensities on 79x0 cards. It shall be fixed in the next version. I'm able to run my 7970s at TCs of 22392 now.
And how much faster is it? :-p 7970 @ 1135/1890, LG 2, TC 22392: GPU 0: 72.0C 3413RPM | 714.6K/715.7Kh/s | A:0 R:1 HW:0 U:0.00/m I:20
No, the memory speed is not a typo. Ignore the accept/reject ratio; I was trying solo. EDIT: The code is now in git, and here's a windows build with the changes: http://ck.kolivas.org/apps/cgminer/temp/cgminer.exe |
|
|
|
|
For all you crazy scrypt/LTC mining fanatics, I have finally found the reason you cannot set very high thread concurrencies or intensities on 79x0 cards. It shall be fixed in the next version. I'm able to run my 7970s at TCs of 22392 now.
|
|
|
|
|
Found a nice awkward bug and the next version of cgminer will fix the high thread concurrencies limitation on 79x0 devices and the intensity 13 limitation.
|
|
|
|
Just a few notes about cgminer that seem to be forgotten with scrypt...
Changing vectors does nothing.
Default lookup gap is 2, so setting 2 or 0 does nothing.
Default gpu threads are 2, so setting 2 does nothing.
Setting thread concurrency overrides shaders, so if you set thread concurrency, setting shaders does nothing.
Intensities over 13 for 79x0 just drops the hashrate or generates garbage so don't do that.
Magic core:mem ratios seem to be entirely hardware dependent: motherboard/CPU/GPU combos change it even for the same 79x0 device.
Magic core:mem ratios work much better on windows. On linux they don't seem to be anywhere near as effective.
Running CPU frequency scaling on linux (or the equivalent on windows) will slow down your GPU hashrate a little, so set your CPU to maximum performance equivalent.
If you want a video, me setting vectors 1 to 2 raising my 7870 to 410 instead of being on 350, without hw errors, let me know. Vectors indeed make a difference, sorry. Considering I wrote cgminer, and that the vectors value isn't even passed to the kernel, I think I'd know. |
|
|
|
|
Driver likely broken. Don't use anything after 12.8.
|
|
|
|
Current one's a bargain at only $285 currently! Get em while they're cheap  |
|
|
|
|
Just a few notes about cgminer that seem to be forgotten with scrypt...
Changing vectors does nothing.
Default lookup gap is 2, so setting 2 or 0 does nothing.
Default gpu threads are 2, so setting 2 does nothing.
Setting thread concurrency overrides shaders, so if you set thread concurrency, setting shaders does nothing.
Intensities over 13 for 79x0 just drops the hashrate or generates garbage so don't do that.
Magic core:mem ratios seem to be entirely hardware dependent: motherboard/CPU/GPU combos change it even for the same 79x0 device.
Magic core:mem ratios work much better on windows. On linux they don't seem to be anywhere near as effective.
Running CPU frequency scaling on linux (or the equivalent on windows) will slow down your GPU hashrate a little, so set your CPU to maximum performance equivalent.
|
|
|
|
[2013-03-14 21:31:47] This error says the device is not enabled
Any pointers appreciated.
sudo aticonfig --adapter=all -f --initial ck, add this command to readme... I see this every 5th page or so ;] I was hoping by now I could have said something like "GPU mining is deprecated, only ASIC code is supported from here on"... but it seems that was a pipedream  , and scrypt is coming back to haunt me with a vengeance  I haven't seen so many new questions about GPU mining in over a year... |
|
|
|
[2013-03-14 21:31:47] This error says the device is not enabled
Any pointers appreciated.
sudo aticonfig --adapter=all -f --initial |
|
|
|
|
Intensities above 9 are a mistake.
|
|
|
|
Note I also improved on the kernel slightly too in the latest version.
Any thoughts on whether a win64 build would improve hash rates? Quite sure it will make absolutely no difference. |
|
|
|
I think the most valuable information is in the caveat from crazyates. My list of totally unscientific, unproven, and completely biased recommendations of GPU brands.
You will only get opinions from people here. There is not much to objectively tell them apart, and some people have really good experiences with one brand while others have the exact opposite experience, and much is left to hearsay. Like, I hear a lot of people find their XFX devices don't allow much change in clocking or cut corners, while others say their warranty is great, double fans are good etc. My own opinion: steer clear of XFX and HIS and maybe diamond. The rest seem okay. I exclusively buy reference model Sapphire because they seem reliably good while being cheaper than others, though I also had a rig where reference designs worked best for cooling. |
|
|
|
I've upgraded to the Catalyst 13.2 beta and now works fine...
Many had problems with Catalyst 13.2 beta and had to go back to 13.1. Maybe there is a new beta that works, though, I don't know. Actually it was the other way around. 13.1 was a steaming pile breaking stuff all over the place whereas 13.2beta+ started working again. |
|
|
|
|
Note I also improved on the kernel slightly too in the latest version.
|
|
|
|
Thanks for clearing this up. It makes a whole lot more sense now. I'll tip you when I find your address.
No tip needed, I have a couple coins already. Someone once told me to truly understand something you need to be able to explain it to someone else. But did that someone explain it to you well? |
|
|
|
Silly question but did you try rebooting after saving your settings? Some things on Linux don't get applied until the boot scripts run. Can be a bit of a hassle if you underclock it enough to stop it rebooting though. You cannot directly software undervolt 7970s on linux. |
|
|
|
|
Driver version is critical. AMD regularly break things.
|
|
|
|
|
Show Posts
