shitty amdgpu-pro mining performance

Wolf0

member

Activity: 81

Merit: 1002

It was only the wind.

Quote from: lhl on November 19, 2016, 11:07:13 PM

So I'm running 16.30 package w/ the 4.8.4-1-ARCH kernel (on arch) on a 6 x RX470 (Hynix) system w/ and I can do a few things w/ it...

I'm currently running a stock ROM.

I can use the overdrive w/o a problem, but pp_sclk_od only seems to affect the top speed, not the intermediate levels, so it's not really useful for me (I underclock for better perf/W and run at P6).

I can also overdrive the pp_mclk_od which would be more useful for me, but the 2000 strap timings are worse than the 1750 strap timings, so it's not a big win. Is there any way to adjust memory timings in Linux w/o booting into Windows/flashing a new VBIOS?

Also, is there anything available for adjusting voltage - really, just voltage offset would be fine (defaults to +4 on my Nitro cards, would ideally just like to bring them down to -10 or -12).

EthOS is using a tool I wrote for them that allows them to set the clocks for any DPM state you want, among other things. You might want to try it.

As for voltage - dude, you obviously know it's set to 4 (+25mV)... so just change it, duh. It's a signed byte. I often use -50mV or -37.5mV on more fussy cards. The only pain in the ass becomes when the VBIOS does NOT have the voltage ID needed for that... then you must add it, and it's a big pain in the ass.

lhl

newbie

Activity: 3

Merit: 0

So I'm running 16.30 package w/ the 4.8.4-1-ARCH kernel (on arch) on a 6 x RX470 (Hynix) system w/ and I can do a few things w/ it...

I'm currently running a stock ROM.

I can use the overdrive w/o a problem, but pp_sclk_od only seems to affect the top speed, not the intermediate levels, so it's not really useful for me (I underclock for better perf/W and run at P6).

I can also overdrive the pp_mclk_od which would be more useful for me, but the 2000 strap timings are worse than the 1750 strap timings, so it's not a big win. Is there any way to adjust memory timings in Linux w/o booting into Windows/flashing a new VBIOS?

Also, is there anything available for adjusting voltage - really, just voltage offset would be fine (defaults to +4 on my Nitro cards, would ideally just like to bring them down to -10 or -12).

smaxz

sr. member

Activity: 430

Merit: 253

VeganAcademy

Quote from: nerdralph on November 19, 2016, 11:42:21 AM

Quote from: smaxz on November 19, 2016, 08:54:14 AM

no experience with R9's just my RX cards that have worked fine.

What kernel version are you using?

4.4.0-47

nerdralph

sr. member

Activity: 588

Merit: 251

Quote from: smaxz on November 19, 2016, 08:54:14 AM

no experience with R9's just my RX cards that have worked fine.

What kernel version are you using?

nerdralph

sr. member

Activity: 588

Merit: 251

Quote from: bobben2 on November 18, 2016, 04:18:34 PM

You might want to try to be root, then run the command to change dpm state.
i.e. sudo is not enough.

The file permissions don't allow anyone but root to change them.

I'm thinking my problem may be related to the kernel. Ubuntu 16.04.1 includes kernel 4.4.0, but overdrive control seems to require at least 4.8:
http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-OverDrive-R9-Fury

Unless someone has a different suggestion, I'm going to try 4.8.9.

smaxz

sr. member

Activity: 430

Merit: 253

VeganAcademy

Quote from: ?? on ??

Quote from: bobben2 on November 18, 2016, 04:18:34 PM

You might want to try to be root, then run the command to change dpm state.
i.e. sudo is not enough.

Of course it's not, not unless you do sudo sh -c 'echo manual > ...'

The reason being the redirect by the shell is run as you, not under sudo.

even as root you will run in to some bugs on having settings take. generally things work though.

no experience with R9's just my RX cards that have worked fine.

isn't AMDGPU-PRO available for ubuntu 14.04 now as well?

bobben2

full member

Activity: 279

Merit: 104

You might want to try to be root, then run the command to change dpm state.
i.e. sudo is not enough.

nerdralph

sr. member

Activity: 588

Merit: 251

I'm still having some problems with the first card, either r9 380 or Rx 470, where it stays in dpm 0 state. I'd like to try the amdgpu-pro 16.30 drivers instead of the 16.40. Anyone know where I could find the old version?

nerdralph

sr. member

Activity: 588

Merit: 251

Quote from: hypercrypto on November 17, 2016, 06:46:28 PM

Quote from: nerdralph on November 17, 2016, 12:14:22 PM

Quote from: nerdralph on November 17, 2016, 10:28:36 AM

It seems the problem is the dpm clock control for the primary video card. Setting power_dpm_force_performance_level to manual and then setting pp_dpm_sclk has no effect (it stays at 0).
https://www.phoronix.com/forums/forum/linux-graphics-x-org-drivers/open-source-amd-linux/871595-amd-delivers-overdrive-overclocking-support-for-amdgpu-drm-driver/page4

I think something with Xorg may be interfering with the clock settings, so I'm going to try running X off the on-board video.

So far trying to run the iGPU along with amdgpu-pro is just as problematic as it was with fglrx. With the iGPU enabled and a Rx 470 and R9 380 plugged in, opencl crashes.

Code:

$ /opt/amdgpu-pro/bin/clinfo
Segmentation fault (core dumped)

try sudo command ,i think it will work

Thanks! That did it, and reminded me about the video group requirement from the driver install directions. I added my mining userid to the video group, and now OpenCL is working. I've left dpm_force_performance to auto for now, as the driver bumps both cards to dpm7 once they start mining. I tried manual to test things out, and for some reason it was only working for mlck. When I 'echo 7 > pp_dpm_sclk' it stayed at 0.

hypercrypto

newbie

Activity: 19

Merit: 0

Quote from: nerdralph on November 17, 2016, 12:14:22 PM

Quote from: nerdralph on November 17, 2016, 10:28:36 AM

It seems the problem is the dpm clock control for the primary video card. Setting power_dpm_force_performance_level to manual and then setting pp_dpm_sclk has no effect (it stays at 0).
https://www.phoronix.com/forums/forum/linux-graphics-x-org-drivers/open-source-amd-linux/871595-amd-delivers-overdrive-overclocking-support-for-amdgpu-drm-driver/page4

I think something with Xorg may be interfering with the clock settings, so I'm going to try running X off the on-board video.

So far trying to run the iGPU along with amdgpu-pro is just as problematic as it was with fglrx. With the iGPU enabled and a Rx 470 and R9 380 plugged in, opencl crashes.

Code:

$ /opt/amdgpu-pro/bin/clinfo
Segmentation fault (core dumped)

try sudo command ,i think it will work

nerdralph

sr. member

Activity: 588

Merit: 251

Quote from: nerdralph on November 17, 2016, 10:28:36 AM

It seems the problem is the dpm clock control for the primary video card. Setting power_dpm_force_performance_level to manual and then setting pp_dpm_sclk has no effect (it stays at 0).
https://www.phoronix.com/forums/forum/linux-graphics-x-org-drivers/open-source-amd-linux/871595-amd-delivers-overdrive-overclocking-support-for-amdgpu-drm-driver/page4

I think something with Xorg may be interfering with the clock settings, so I'm going to try running X off the on-board video.

So far trying to run the iGPU along with amdgpu-pro is just as problematic as it was with fglrx. With the iGPU enabled and a Rx 470 and R9 380 plugged in, opencl crashes.

Code:

$ /opt/amdgpu-pro/bin/clinfo
Segmentation fault (core dumped)

Subw

hero member

Activity: 672

Merit: 500

can't speak about R9 380 as i don't have any but R9 390 hashes OK both with old fglrx (ubuntu 15) and new amdpro (ubuntu 16) drivers (about 90-100S/s with recent SA miner)

bobben2

full member

Activity: 279

Merit: 104

I gave up equihash mining with my Tongas since they tended to crash the miner.
I suspect mem clock too high (1625 - too lazy to bios mod).
They hash ok when mining Eth (23.1MH/s) both with fglrx and amdgpu-pro driver (16.30)

I tried optiminer 0.3.3, but that halfed the hashrate, so I reverted back to 0.3.2. This with amdgpu-pro driver

nerdralph

sr. member

Activity: 588

Merit: 251

It seems the problem is the dpm clock control for the primary video card. Setting power_dpm_force_performance_level to manual and then setting pp_dpm_sclk has no effect (it stays at 0).
https://www.phoronix.com/forums/forum/linux-graphics-x-org-drivers/open-source-amd-linux/871595-amd-delivers-overdrive-overclocking-support-for-amdgpu-drm-driver/page4

I think something with Xorg may be interfering with the clock settings, so I'm going to try running X off the on-board video.

nerdralph

sr. member

Activity: 588

Merit: 251

Until this week I've been using fglrx on Ubuntu 14.04 for mining. Yesterday I tried upgrading a 14.04 rig to 16.04 and the amdgpu-pro drivers. I was running into a lot of problems, so today I did a fresh 16.04.1 (4.4.0-31 kernel) install with just 2 cards installed; a R9 380 and a Rx 470. I used the latest 16.40 amdgpu-pro drivers. The performance on the R9 380 sucks; sa-v5 gives me <10 sols/s, while the same card gives me ~50 with fglrx. The Rx 470 is better (60-65 sols).

Anyone know what is required to get decent performance for Tonga with amdgpu-pro? Or should I only use it for Ellesmere and stick with fglrx for everything else?

Topic: shitty amdgpu-pro mining performance (Read 3198 times)