Examples of AMD Gpu commandlines - FAQ

Message boards : Number crunching : Examples of AMD Gpu commandlines - FAQ
Message board moderation

To post messages, you must log in.

1 · 2 · Next

AuthorMessage
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1993920 - Posted: 16 May 2019, 2:27:36 UTC
Last modified: 16 May 2019, 2:36:34 UTC

My one AMD GPU is currently down (its an APU) but it seems like it might be useful to consolidate as much as we can, all the command line advice for AMD video cards (from the top to the bottom). So here are some results in no particular order. We need some good ideas for mid/entry level gpus too.

Here is Mike's response to a Vega 56 question in the "GPU FLOPS: Theory vs Reality" thread.


https://setiathome.berkeley.edu/forum_thread.php?id=81962&postid=1963448#1963448

A good start for Vega would be.

-sbs 2048 -period_iterations_num 1 -spike_fft_thresh 4096 -high_perf -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64


---------------------------
Here is Mr. Rueike's (sp) offering for Vega 64/Navi VII gpus.
https://setiathome.berkeley.edu/forum_thread.php?id=84174&postid=1992634#1992634
-v 1 -instances_per_device 1 -sbs 2048 -period_iterations_num 1 -tt 600 -spike_fft_thresh 4096 -oclfft_tune_gr 2<span class="mark">56</span> -oclfft_tune_lr 16 -oclfft_tune_wg 2<span class="mark">56</span> -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 -hp -high_perf -no_defaults_scaling -cpu_lock -nobs -tune 1 64 1 4 -no_use_sleep


Tom
A proud member of the OFA (Old Farts Association).
ID: 1993920 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 1993922 - Posted: 16 May 2019, 2:33:42 UTC

Rick's command line is more appropriate for the top of the AMD product line. As evidenced by his best of all Seti AMD card's computation times, I think the use of the -hp -high_perf -no_defaults_scaling -cpu_lock -nobs -tune 1 64 1 4 -no_use_sleep tuning parameters, particularly the -nobs and the -cpu_lock make the biggest difference.

Mike's more generic command line parameters are a good compromise for the entire AMD product family.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 1993922 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1993941 - Posted: 16 May 2019, 11:52:22 UTC - in response to Message 1993922.  

For Windows based machines the command-line files are in the hidden directory/sub-directory "ProgramData\BOINC\Projects\setiathome.berkely.edu".

You can find this by toggling the "hidden files" in the File Explorer. In Win10 that is under the "view".

The command files for MB/AP are all .txt files. The naming convention(s) are usually exactly/very similar to the names of the applications that they provide command lines for.

AMD gpus have 3 versions of the same gpu application. So you need to put the same parameters into 3 slightly differently named command line files.

The notepad utility in Windows is a bit cranky when you want to create/modify files that don't end in .txt (like app_config.xml). You need to create the file and save it as an "all files" file. If you make a mistake you will end up with a file named app_config.xml.txt which doesn't work. You have to rename it. One trick is to set the File Explorer to always display extensions.

As far as I know, if you are running the standard version(s) of BOINC under linux, the same commandline parameters apply.
A proud member of the OFA (Old Farts Association).
ID: 1993941 · Report as offensive
Richard Haselgrove Project Donor
Volunteer tester

Send message
Joined: 4 Jul 99
Posts: 14649
Credit: 200,643,578
RAC: 874
United Kingdom
Message 1993944 - Posted: 16 May 2019, 12:08:05 UTC - in response to Message 1993941.  

Hidden "by default". Users who want to fiddle around with configurations have usually chosen a more convenient (and visible) data location during installation, preferably not on their SSD system drive.
ID: 1993944 · Report as offensive
Profile Jaye Ellen

Send message
Joined: 29 Nov 08
Posts: 26
Credit: 20,945,032
RAC: 45
United States
Message 1993961 - Posted: 16 May 2019, 17:31:06 UTC - in response to Message 1993941.  

Is this the same Tom that was helping me before? if so, I need your help again, if you're available!
ID: 1993961 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1993977 - Posted: 16 May 2019, 22:58:29 UTC - in response to Message 1993961.  

Is this the same Tom that was helping me before? if so, I need your help again, if you're available!


Yes I am Jaye Ellen. PM.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1993977 · Report as offensive
Profile StFreddy
Avatar

Send message
Joined: 4 Feb 01
Posts: 35
Credit: 14,080,356
RAC: 26
Hungary
Message 1994013 - Posted: 17 May 2019, 15:25:11 UTC

I used both Mike's and Rueike's command line parameters on my undervolted Vega56. Not much difference, just few seconds. OS: windows 10.
Based on my findings, a Ryzen 1800x @3.9GHz (1.25V) is much more efficient than my Vega 56: roughly the same RAC with lower power consumption (blc33):
- Vega56: ~19k RAC/day, ~150W power consumption
- Ryzen 1800x: ~19k RAC/day , ~105W power consumption

this is under windows 10 with the latest Lunatics Apps installed. This is why I won't use this Vega for Seti any more.
Maybe I will find another project for Vega where it may be more efficient.
ID: 1994013 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1994020 - Posted: 17 May 2019, 15:47:34 UTC - in response to Message 1994013.  

I used both Mike's and Rueike's command line parameters on my undervolted Vega56. Not much difference, just few seconds. OS: windows 10.
Based on my findings, a Ryzen 1800x @3.9GHz (1.25V) is much more efficient than my Vega 56: roughly the same RAC with lower power consumption (blc33):
- Vega56: ~19k RAC/day, ~150W power consumption
- Ryzen 1800x: ~19k RAC/day , ~105W power consumption

this is under windows 10 with the latest Lunatics Apps installed. This is why I won't use this Vega for Seti any more.
Maybe I will find another project for Vega where it may be more efficient.


I understand. And everything I have read/experienced supports the point that NVIDIA gpus usually use less power than the equivalent AMD gpus. I am hopeful that will change as the new Navi gpu's roll out.

Since we have new people who roll in here with AMD gpus it seems reasonable to offer them the best commandline advice we can.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1994020 · Report as offensive
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 13161
Credit: 1,160,866,277
RAC: 1,873
United States
Message 1994027 - Posted: 17 May 2019, 16:17:11 UTC - in response to Message 1994020.  

All the rumors hint that there will be no changes in power consumption for Navi compared to previous generations. Still power hogs.
Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)
ID: 1994027 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 1994048 - Posted: 17 May 2019, 18:04:35 UTC

A great thread to start up, thanks, Tom! Refresh my memory, are these command lines applicable to either stock apps or Lunatics? Do they need to be applied to MB and AP separately?

I have a Vega 8 GPU (Ryzen 3 2200G APU), so would Mike's command line be a good one to start off with?

I think I am having a similar situation as StFreddy. It seems that my CPU tasks complete in about the same time, if not less, than the Vega 8 GPU. I'm sure the GPU is better suited for a different project, but I'm curious to see what other tweaking I can do to it first before switching it over.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 1994048 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1994074 - Posted: 17 May 2019, 20:46:00 UTC - in response to Message 1994048.  
Last modified: 17 May 2019, 20:56:54 UTC

A great thread to start up, thanks, Tom! Refresh my memory, are these command lines applicable to either stock apps or Lunatics? Do they need to be applied to MB and AP separately?

I have a Vega 8 GPU (Ryzen 3 2200G APU), so would Mike's command line be a good one to start off with?

I think I am having a similar situation as StFreddy. It seems that my CPU tasks complete in about the same time, if not less, than the Vega 8 GPU. I'm sure the GPU is better suited for a different project, but I'm curious to see what other tweaking I can do to it first before switching it over.


There is a trick that is specific to the 2200G/2400G. 1) Only allow 75% of the cpu's to run (BOINC Manager, local). 2) Dedicate 1 cpu to the gpu. (app_config.xml) 3) Run the ram as fast as you can and still be stable. 4) Turn up the frequency of the iGPU to someplace in the ballpark of 1500(?) for the 2400G, not sure what it is for the 2200G.

These command lines will work on any version of the Windows gpu BOINC/Seti apps. I have used them with the Lunatics distro. This command line is for the MB app. The AP app has a different command line and/or command line file. And the parameters recommended are different.

They will also work on the "stock" aka: repository Linux version of BOINC/Seti.

They should work on the Tbar All-in-One version of Linux/BOINC except you will have to manually add the AMD apps (and download them) to the "app_info.xml" file.

I have been running a Ryzen 5 2400G with iGPU. And the gpu was consistently processing at least twice as fast as the CPU tasks were. Unfortunately, I have had bad luck and it refused to boot a couple of weeks ago. If I remember right, my CPU tasks were running 2.5+ hours and my gpu tasks were always under an hour and a half. Frequently they took about an hour. Sometimes 47 minutes.

The Windows readme files for AMD gpu's include an example of a low-end card that has smaller parameter settings than Mike's generic starting command line.
I can say that setting -tt 1500 and -period_iterations_num 1 did not cause my iGPU to "lag".

HTH,
Tom
A proud member of the OFA (Old Farts Association).
ID: 1994074 · Report as offensive
Profile Mike Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 17 Feb 01
Posts: 34253
Credit: 79,922,639
RAC: 80
Germany
Message 1994075 - Posted: 17 May 2019, 20:52:59 UTC

First of all, each command line is better than no command line.
As StFreddy mentioned the difference between my values and Rick´s are just a few seconds.
That`s the reason i wrote testing required in the readme.
It makes a difference only if one is running multiple GPU`s.

My 1800X is also faster than my R9 380 but thats no surprise.

Bill, you can`t compare a Vega 8 against a Vega 56.
You would benefit more when using optimized CPU apps.


With each crime and every kindness we birth our future.
ID: 1994075 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 1994136 - Posted: 18 May 2019, 2:03:08 UTC - in response to Message 1994074.  

There is a trick that is specific to the 2200G/2400G. 1) Only allow 75% of the cpu's to run (BOINC Manager, local). 2) Dedicate 1 cpu to the gpu. (app_config.xml) 3) Run the ram as fast as you can and still be stable. 4) Turn up the frequency of the iGPU to someplace in the ballpark of 1500(?) for the 2400G, not sure what it is for the 2200G.

1. I have more to say about this later, but for now, since the 2200G is a 4 core 4 thread processor, does that matter? I thought I read here once that scaling back on CPUs make a difference because of multi threaded processors.
2. Done already.
3. I have my XMP profile set for 3200, which is the speed of the RAM I bought. I have started looking at 1usmus' DRAM calculator, but I haven't changed anything yet. After reading up on Gamers' Nexus that the SOC voltage read on your computer may be different that what is physically powered, I figured it would be best that I read through the steps thoroughly and be sure I know what I'm doing. Keith posted an article on Einstein's board which is really good, so I'll post it again here.
4. Combining this with #1 above, I have questions about this. I've started a new thread so I don't derail the conversation about command lines. Check it out here.

First of all, each command line is better than no command line.
Okay, I was not sure of that. I didn't know if entering a command line that wasn't "compatible" with your specific processor would cause problems other than it possibly slowing down.

Bill, you can`t compare a Vega 8 against a Vega 56.
I'm not sure I understand what you mean. I was comparing that the GPU tasks take about the same time, if not more in some cases, than a CPU task. The only comparison I was trying to make is that perhaps my Vega 8 would get more work done crunching a different application or project than what it is doing now. I don't know that I realized this until recently.

You would benefit more when using optimized CPU apps.
Yeah, I know you are right. I have been of the opinion of being a "stock app purist", but I am starting to change my view on that. However, there is still a problem with the peak GFlops error in the stock AP7 tasks that hasn't been patched yet, and I want to at least see that through (holding my breath here).
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 1994136 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1996035 - Posted: 30 May 2019, 23:06:12 UTC

If you are running an AMD 3 2200G or an AMD 5 2400G and you are processing tasks with the iGPU what kind of command line are you using?

Tom
A proud member of the OFA (Old Farts Association).
ID: 1996035 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 1996074 - Posted: 31 May 2019, 12:09:35 UTC - in response to Message 1996035.  

If you are running an AMD 3 2200G or an AMD 5 2400G and you are processing tasks with the iGPU what kind of command line are you using?

Tom
I haven't gotten around to adding a command line yet, but I was thinking of adding the first Vega line you put in your first post.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 1996074 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1996113 - Posted: 31 May 2019, 19:44:37 UTC - in response to Message 1996074.  
Last modified: 31 May 2019, 20:07:57 UTC

If you are running an AMD 3 2200G or an AMD 5 2400G and you are processing tasks with the iGPU what kind of command line are you using?

Tom
I haven't gotten around to adding a command line yet, but I was thinking of adding the first Vega line you put in your first post.


I am playing with some and so far I haven't seen a speedup or even a slow down :(

So after toggling some different things to try to get my iGPU to run at 1600 I have stripped my command line to the "entry level" command line.


-spike_fft_thresh 2048 -tune 1 2 1 16 


To see if that makes ANY difference.

I am getting huge wallclock times with only 3 odd minutes of cpu time. So I have been wondering how to "reduce" the wallclock time.
-edit-
I was looking at your tasks and your cpu is running them in a bit over an hour. Mine are taking 1 3/4 hours. Your gpu is taking maybe 1 hour 20 minutes. Mine is taking 49 minutes up to an hour 1 15 plus.
-edit-
Makes me wonder if my SMT is slowing things down that much on the cpu side.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1996113 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1996191 - Posted: 1 Jun 2019, 2:53:48 UTC - in response to Message 1996113.  

Added -sbs 1024

Tried the "midlevel" equivalent of my command line and watched the load on the gpu (under the Windows task manager) appear to decline a couple of %. So back to the entry level. Will try running some of the "other parameters" in and out to see if anything causes the gpu to peg to 100% more firmly. It certainly isn't now.

Still not sure why my cpu times are so slow compared to yours.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1996191 · Report as offensive
Profile Bill Special Project $75 donor
Volunteer tester
Avatar

Send message
Joined: 30 Nov 05
Posts: 282
Credit: 6,916,194
RAC: 60
United States
Message 1996418 - Posted: 2 Jun 2019, 13:15:17 UTC - in response to Message 1996113.  

I am getting huge wallclock times with only 3 odd minutes of cpu time. So I have been wondering how to "reduce" the wallclock time.
-edit-
I was looking at your tasks and your cpu is running them in a bit over an hour. Mine are taking 1 3/4 hours. Your gpu is taking maybe 1 hour 20 minutes. Mine is taking 49 minutes up to an hour 1 15 plus.
-edit-
Makes me wonder if my SMT is slowing things down that much on the cpu side.

Tom
Yeah, we have encountered before that your 2400G seems to perform slower than my 2200G, and I don't think we ever got to the bottom of it. Yes, you have SMT available, whereas it isn't available for my processor. Just curious, how fast are you running your ram? Number of DIMMs? And I'm assuming they're in the optimal slots? I don't want to hijack this thread, so perhaps we continue this portion of the conversation somewhere else.

I put in my command line the following (from the first post):

-sbs 2048 -period_iterations_num 1 -spike_fft_thresh 4096 -high_perf -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64
And I don't think it has really changed the performance. If anything, the may have increased by a few minutes. I'll have to take a look at this closer.
Seti@home classic: 1,456 results, 1.613 years CPU time
ID: 1996418 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1996460 - Posted: 2 Jun 2019, 19:45:09 UTC - in response to Message 1996418.  

I am getting huge wallclock times with only 3 odd minutes of cpu time. So I have been wondering how to "reduce" the wallclock time.
-edit-
I was looking at your tasks and your cpu is running them in a bit over an hour. Mine are taking 1 3/4 hours. Your gpu is taking maybe 1 hour 20 minutes. Mine is taking 49 minutes up to an hour 1 15 plus.
-edit-
Makes me wonder if my SMT is slowing things down that much on the cpu side.

Tom
Yeah, we have encountered before that your 2400G seems to perform slower than my 2200G, and I don't think we ever got to the bottom of it. Yes, you have SMT available, whereas it isn't available for my processor. Just curious, how fast are you running your ram? Number of DIMMs? And I'm assuming they're in the optimal slots? I don't want to hijack this thread, so perhaps we continue this portion of the conversation somewhere else.

I put in my command line the following (from the first post):

-sbs 2048 -period_iterations_num 1 -spike_fft_thresh 4096 -high_perf -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64
And I don't think it has really changed the performance. If anything, the may have increased by a few minutes. I'll have to take a look at this closer.


I am trying this one right now:

-sbs 1024 -period_iterations_num 1 -spike_fft_thresh 2048 -high_perf -tune 1 2 1 16 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 -cpu_lock


It has two differences. A smaller -sbs and the -tune 1 2 1 16 which is for an "entry level" video card. Short term testing seemed to indicate that the smaller -tune kept the video card busier.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1996460 · Report as offensive
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 5124
Credit: 276,046,078
RAC: 462
Message 1996495 - Posted: 3 Jun 2019, 0:52:24 UTC - in response to Message 1996460.  

I am getting huge wallclock times with only 3 odd minutes of cpu time. So I have been wondering how to "reduce" the wallclock time.
-edit-
I was looking at your tasks and your cpu is running them in a bit over an hour. Mine are taking 1 3/4 hours. Your gpu is taking maybe 1 hour 20 minutes. Mine is taking 49 minutes up to an hour 1 15 plus.
-edit-
Makes me wonder if my SMT is slowing things down that much on the cpu side.

Tom
Yeah, we have encountered before that your 2400G seems to perform slower than my 2200G, and I don't think we ever got to the bottom of it. Yes, you have SMT available, whereas it isn't available for my processor. Just curious, how fast are you running your ram? Number of DIMMs? And I'm assuming they're in the optimal slots? I don't want to hijack this thread, so perhaps we continue this portion of the conversation somewhere else.

I put in my command line the following (from the first post):

-sbs 2048 -period_iterations_num 1 -spike_fft_thresh 4096 -high_perf -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64
And I don't think it has really changed the performance. If anything, the may have increased by a few minutes. I'll have to take a look at this closer.


I am trying this one right now:

-sbs 1024 -period_iterations_num 1 -spike_fft_thresh 2048 -high_perf -tune 1 2 1 16 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64 -cpu_lock


It has two differences. A smaller -sbs and the -tune 1 2 1 16 which is for an "entry level" video card. Short term testing seemed to indicate that the smaller -tune kept the video card busier.

Tom


I missed that I also made the -spike_fft_thresh smaller.

Tom
A proud member of the OFA (Old Farts Association).
ID: 1996495 · Report as offensive
1 · 2 · Next

Message boards : Number crunching : Examples of AMD Gpu commandlines - FAQ


 
©2024 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.