Rather than making separate posts in each project, this thread will serve to document the performance of various AMD GPU projects running on Linux with the new AMDGPU-Pro drivers.
The first round of tests was run in Ubuntu 16.04 LTS with an AMD RX 480 GPU. For the Windows results, I ran the GPU on Windows 10 with the 16.7.2 drivers. Depending on the performance of the RX 480, I may also add some data for the R9 290 and the HD 7970. The 290 and 7970 will probably not be running in Ubuntu, but rather, they will be running in Arch Linux with the latest kernel patched for the AMDGPU-Pro drivers. The results so far are promising and should be even better as the Linux driver continues to mature.
The projects that I will be running are: POEM@Home, Moo!, Collatz Conjecture, PrimeGrid, Einstein@Home, and SETI@Home.
Update 1/31/17
The setup for round two is a little different, and I hope to incorporate more results than I had in Round 1 of testing. I will also set the GPU clocks in Windows equal to Linux, so that it is a more straightforward comparison.
For the second round of testing I will be running Arch Linux with the 4.9 kernel, and the AMD 16.60 driver. For the Windows results, I will be running Windows 10 Pro. I'm not 100% sure which driver I will run, but it will probably be the latest driver.
For the PPD numbers for each project, I ran each project for a minimum of one day, and then randomly selected 60 completed and validated tasks to use for the performance summary.
app_config.xml: N/A - No app_config used for this project
No project specific configurations were used
GPU memory clocks were frequently dropping to idle, but looking at other system reults on Moo!, this does not appear to affect performance.
Force memory clock to run at full speed: After running 4 tasks with memory clocks forced to full speed, there was no noticable difference in task runtimes.
Einstein@Home
app_config.xml: N/A - No app_config used for this project
All Einstein configurations were done via the project preferences on my account page at Einstein@Home
Run Test Applications -> Yes
GPU utilization factor of GRP,FGRP, and GW apps -> .5
MilkyWay@Home
app_config.xml: Simple app_config to run two tasks concurrently
I can do some power consumption numbers. It will just be the measured system power at the UPS. I'll do some idle and load numbers for Windows and Ubuntu.
Sounds great, and many thanks! I might look into one of these per dual Xeon setup that I'm planning since I really like those numbers, and when gaming they don't seem to pull that much power... Oh, and Watercool makes a Heatkiller block for it (or is in the process of doing such) while most companies seem to be ignoring the more expensive GTX 1060 for blocks (with the 1060 being upwards of $100 more than the RX 480 up here).
My biggest curiosity is the numbers themselves.... Power consumption data was just icing on the cake as I'm seriously contemplating the idea of two Natex bundles, an L5640 in the X8DTT, and then a 5960X...... Along with three high end GPU's ^_^;;;;
Here are some overall numbers from a few projects. The GPU was OC'd to 1355 core 2150 memory for all the Windows tasks. To normalize the results subtract 7% or so from the Windows results, and that should give a fairly accurate representation of the stock Windows performance. As soon as I move the 480 to it's permanent home under water I will flash the BIOS and then see how it performs in Linux with an overclock.
POEM@Home
Linux 1266/2000
Points
Average Time (sec)
PPD
9600
1556
532945
9100
1706
460868
5500
674
705393
Overall PPD = 566,402
Windows 1355/2150
Points
Average Time (sec)
PPD
9600
1585
523471
9100
1353
580966
5500
659
720756
Overall PPD = 608,398
Collatz Conjecture
Linux 1266/2000
Average Points
Average Time (sec)
PPD
3706
116
2760213
Overall PPD = 2,760,213
Windows 1355/2150
Average Points
Average Time (sec)
PPD
3881
107
3130232
Overall PPD = 3,130,232
Moo! Wrapper
Linux 1266/2000
Points
Average Time (sec)
PPD
6584
971
585948
6144
908
584894
368
70
453246
256
53
419038
Overall PPD = 510,781
Windows 1355/2150
Points
Average Time (sec)
PPD
6144
727
730543
Overall PPD = 730,543
* Moo! only ran the smaller tasks when the project was first added, after that it only ran the larger tasks. I haven't had time to sort out the big difference in PPD on this project. It may be that there are some optimizations in OpenCL 2.0, that Moo! is taking advantage of. (current AMD Linux driver is built with OpenCL 1.2)
I ended up moving the RX 480 into a Windows machine, so it looks like the RX 480 Linux testing will have to be put on hold. Once the next major AMD Linux driver update comes out, I will try to move the 480 back into a Linux machine.
I did, but it was right before I went out of town, and those results have been wiped from my history on Einstein.
I saw your post in the other thread, so I just loaded up some Einstein tasks.
Normally I would run 3 at a time, but I am only going to run a single task to make it easier to parse the data. *Edit* Just changed it to 2 at a time, so I get some realistic PPD numbers. I will let it roll at stock clocks, and see what it does over the next couple of days.
I will be doing a new comparison with the release of the latest Linux driver. I am running Folding@Home on the 480 with the latest driver right now, and so far it looks like there is a roughly 20% improvement with the new driver.
Hopefully I will have time to do a pretty thorough comparison. It will take roughly a week to run through all the projects in Linux and Windows. I think I will get started on Sunday, so stay tuned for results.
For the Linux results, I will be running Arch Linux since that is what I run on my test server. What flavor of Windows would everyone like to see for the comparison?
I will be doing a new comparison with the release of the latest Linux driver. I am running Folding@Home on the 480 with the latest driver right now, and so far it looks like there is a roughly 20% improvement with the new driver.
Hopefully I will have time to do a pretty thorough comparison. It will take roughly a week to run through all the projects in Linux and Windows. I think I will get started on Sunday, so stay tuned for results.
For the Linux results, I will be running Arch Linux since that is what I run on my test server. What flavor of Windows would everyone like to see for the comparison?
I think that Windows 10 Home x64 is what most people will be using if they play games on AMD Radeons; most DX12 introduced improvements would only be in Windows 10. My BOINC system is on Windows 7 Pro x64 though, although I once had a Haswell i5 on Windows 8.1 running ATLAS or WCG which functionally is Windows 10's backend.
Most reviews are running Windows 10 now.
EDIT: By the way thanks for the continual contributions to our knowledge of Polaris' BOINC performance.
I will be doing a new comparison with the release of the latest Linux driver. I am running Folding@Home on the 480 with the latest driver right now, and so far it looks like there is a roughly 20% improvement with the new driver.
Hopefully I will have time to do a pretty thorough comparison. It will take roughly a week to run through all the projects in Linux and Windows. I think I will get started on Sunday, so stay tuned for results.
For the Linux results, I will be running Arch Linux since that is what I run on my test server. What flavor of Windows would everyone like to see for the comparison?
I think that Windows 10 Home x64 is what most people will be using if they play games on AMD Radeons; most DX12 introduced improvements would only be in Windows 10. My BOINC system is on Windows 7 Pro x64 though, although I once had a Haswell i5 on Windows 8.1 running ATLAS or WCG which functionally is Windows 10's backend.
Most reviews are running Windows 10 now.
EDIT: By the way thanks for the continual contributions to our knowledge of Polaris' BOINC performance.
Windows 10 is probably the way to go. I will also use the latest ReLive driver for the Windows comparison. In addition to the 480 info, I will also throw up some Linux results for Hawaii and Fiji.
Collatz has seen a bump in Linux performance. The old tests were with the smaller tasks, but I believe that the PPD between the old and new tasks is roughly the same. Running Moo! now and after the first few tasks it also appears to have gotten a boost in performance on Linux.
Once I have a few more results, I will update the thread, and start adding results to the 2nd post.
I just got my RX 480 working in Linux... I've only tried Milkway so far, my times are around 87s. Is that good for a 480? It's definitely better than my gtx 970, and the wingman on my valid tasks using an AMD Tahiti (3072MB) took 115 seconds for the same tasks.
I just got my RX 480 working in Linux... I've only tried Milkway so far, my times are around 87s. Is that good for a 480? It's definitely better than my gtx 970, and the wingman on my valid tasks using an AMD Tahiti (3072MB) took 115 seconds for the same tasks.
That sounds really good. My 7970s clocked at 1200MHz average about 62 seconds per task. Those times are in Windows, since MilkyWay used to not run very well in Linux. I might need to switch that rig over to Linux if the production is that good.
*Edit* I'll run some tasks on my 290s tonight to see how they fare in Linux.
That sounds really good. My 7970s clocked at 1200MHz average about 62 seconds per task. Those times are in Windows, since MilkyWay used to not run very well in Linux. I might need to switch that rig over to Linux if the production is that good.
*Edit* I'll run some tasks on my 290s tonight to see how they fare in Linux.
I don't know what my 480 is clocked. I never looked. It's a XFX RX 480 RS 8GB. I'm happy to finally get it working....had to install Arch to get it to work.
That sounds really good. My 7970s clocked at 1200MHz average about 62 seconds per task. Those times are in Windows, since MilkyWay used to not run very well in Linux. I might need to switch that rig over to Linux if the production is that good.
*Edit* I'll run some tasks on my 290s tonight to see how they fare in Linux.
I don't know what my 480 is clocked. I never looked. It's a XFX RX 480 RS 8GB. I'm happy to finally get it working....had to install Arch to get it to work.
You can check your clocks at /sys/class/drm/card0/device/pp_dpm_sclk for core clocks and pp_dpm_mclk for memory clocks. You can set the fan speed at /sys/class/drm/card0/device/hwmon/hwmon3/pwm1.
Doing that from memory, but I think those are the correct paths. card0 and hwmon3 might be different depending on what other hardware you have installed.
Somehow I think I screwed things up trying to get folding working....which i haven't gotten working yet.... Now the milkway takes are taking 10 seconds longer....unless they are different tasks.
Somehow I think I screwed things up trying to get folding working....which i haven't gotten working yet.... Now the milkway takes are taking 10 seconds longer....unless they are different tasks.
I see 8-9 second variance in run-times, so you are probably OK.
*Edit*@Diffident I forgot to ask if you were running a single task at a time. The run-times I gave you for my 7970s, were for two tasks per GPU which has higher overall throughput.
I am runnning some MilkyWay on my 290s which are clocked at 1000 core, 1350 mem. Running one task at a time the average runtime is 69 seconds. Running two tasks concurrently the average is about 116 seconds.
Would love to know you're einstein results in linux...and if you can actually run more than 2 tasks at a time without the memory controller bugging out....unlike windows
This is an older thread, you may not receive a response, and could be reviving an old thread. Please consider creating a new thread.
Related Threads
?
?
?
?
?
Ask a question
Ask a question
Overclock.net
27.8M posts
541.2K members
Since 2004
A forum community dedicated to overclocking enthusiasts and testing the limits of computing. Come join the discussion about computing, builds, collections, displays, models, styles, scales, specifications, reviews, accessories, classifieds, and more!