Thanks for the update. So has this method worked OK with your RX560?
I have ubuntu 16.04.3 starting from 16.04.1 but kernel has no updated and it is still 4.4 so I think I would have to upgrade to 4.10 before trying this method?
Btw, how many concurrent WUs do you intend to run?
Thanks for the update. So has this method worked OK with your RX560?
I have ubuntu 16.04.3 starting from 16.04.1 but kernel has no updated and it is still 4.4 so I think I would have to upgrade to 4.10 before trying this method?
Btw, how many concurrent WUs do you intend to run?
Yes, this method has worked with my RX560. You can view my list of computers if you wish.
And I would bring Ubuntu up to the latest version.
As for concurrent WUs I am currently running 2 GPUs with a single CPU. Trying to get a feel for where I am with this card.
I have ubuntu 16.04.3 starting from 16.04.1 but kernel has no updated and it is still 4.4 so I think I would have to upgrade to 4.10 before trying this method?
I believe kernel 14.10 has been a requirement since amdgpu-pro 17.30.
Ok two units, so 325-350 Kppd, like a HD7870 more or less. Thanks
I don't think you will be able to get that many credits from an RX 560.
I had a look at robl's tasks list and have made some assumptions about what robl has been doing :-). He'd better correct me quickly if I'm way off the mark :-). It appears that the new GPU started at 2x, returning times (GPU/CPU) around the 2100s/100s mark. The low CPU usage is very nice with AMD cards. There was apparently a change to 3x for a period and the new times were around 3080s/160s. Then, later on, there seems to have been a return to 2x with the times going back to slightly higher than the initial values. This is to be expected because crunch times do increase with the higher frequencies in play for tasks later in a given data series. So, it's not surprising to see an extra minute or so added on for tasks later in a series.
On the basis of those very quick and dirty guesstimates, the theoretical RAC for 2x/3x (GPU contribution only) is around 280K/290K respectively. So, only a very marginal improvement in going to 3x by the look of things. If it were my card I'd do exactly the same (return to 2x) simply because of less likelihood of problems and potentially lower stress/lower power use at the 2x configuration.
I'm grateful to robl for posting the details of this new card - on a number of fronts. I have a lot of RX 460s and also some 560s now. Because of the price differential, they are all of the 2GB variety and so pretty much limited to 2x. I had assumed that there probably wouldn't be much of an increase by going to 3x so was never tempted to pay the extra and buy 4GB cards. Robl's results seem to confirm that view.
Also, I now have RX 460, 560, 570, 580 cards crunching. They are all running on an 'unsupported' linux distro (PCLinuxOS) on OpenCL components extracted from the amdgpu-pro 16.60 package by looking at what the --compute option would install on a 'supported' distro. My 560s are returing a theoretical RAC of around 265K so it looks like (based on robl's figures) there might be a bit on an improvement by looking into extracting components from the 17.50 package. I guess it's time to do a bit of research :-).
... I am completing about 4 to 5 CPU WUs just to get them out of the way. Then i will do 2 concurrent GPU WUs only.
I didn't notice much impact on GPU crunch times when running one or two CPU tasks as well. It does make a bit of difference to power draw though. With the older CPUs (2008 - 2010) that I've brought out of retirement I thought there would be an impact but there doesn't seem to be. I mostly run just one CPU task but on some quads I've been running two.
You can run Einstein @ Home on Ubuntu 16.04 without having to install AMD drivers.The 4.10 kernel already comes with the amdgpu driver, and you just need to install the OpenCL packages: libclc-amdgcn, libclc-dev, libclc-r600, mesa-opencl-icd.
The problem is that you will not be able to run other projects like Milkyway (I haven't yet tested Seti), probably because it is OpenCL 1.1 and not 1.2 like the AMD drivers.
Short update: running E@H full time on Ubu 16 and AMD RX 560 at a 2 GPU WUs concurrency now has me at about 230K avg credit per day. I did allow for a couple days of the new sky WUs to process but have stopped accepting any additional work for them.
Thanks for the update. So has
)
Thanks for the update. So has this method worked OK with your RX560?
I have ubuntu 16.04.3 starting from 16.04.1 but kernel has no updated and it is still 4.4 so I think I would have to upgrade to 4.10 before trying this method?
Btw, how many concurrent WUs do you intend to run?
Trotador wrote:Thanks for the
)
Yes, this method has worked with my RX560. You can view my list of computers if you wish.
And I would bring Ubuntu up to the latest version.
As for concurrent WUs I am currently running 2 GPUs with a single CPU. Trying to get a feel for where I am with this card.
Yes, I went ahead and updated
)
Yes, I went ahead and updated to current kernel.
Ok two units, so 325-350 Kppd, like a HD7870 more or less. Thanks
Trotador wrote:... I have
)
I believe kernel 14.10 has been a requirement since amdgpu-pro 17.30.
Trotador wrote:Yes, I went
)
I don't think you will be able to get that many credits from an RX 560.
I had a look at robl's tasks list and have made some assumptions about what robl has been doing :-). He'd better correct me quickly if I'm way off the mark :-). It appears that the new GPU started at 2x, returning times (GPU/CPU) around the 2100s/100s mark. The low CPU usage is very nice with AMD cards. There was apparently a change to 3x for a period and the new times were around 3080s/160s. Then, later on, there seems to have been a return to 2x with the times going back to slightly higher than the initial values. This is to be expected because crunch times do increase with the higher frequencies in play for tasks later in a given data series. So, it's not surprising to see an extra minute or so added on for tasks later in a series.
On the basis of those very quick and dirty guesstimates, the theoretical RAC for 2x/3x (GPU contribution only) is around 280K/290K respectively. So, only a very marginal improvement in going to 3x by the look of things. If it were my card I'd do exactly the same (return to 2x) simply because of less likelihood of problems and potentially lower stress/lower power use at the 2x configuration.
I'm grateful to robl for posting the details of this new card - on a number of fronts. I have a lot of RX 460s and also some 560s now. Because of the price differential, they are all of the 2GB variety and so pretty much limited to 2x. I had assumed that there probably wouldn't be much of an increase by going to 3x so was never tempted to pay the extra and buy 4GB cards. Robl's results seem to confirm that view.
Also, I now have RX 460, 560, 570, 580 cards crunching. They are all running on an 'unsupported' linux distro (PCLinuxOS) on OpenCL components extracted from the amdgpu-pro 16.60 package by looking at what the --compute option would install on a 'supported' distro. My 560s are returing a theoretical RAC of around 265K so it looks like (based on robl's figures) there might be a bit on an improvement by looking into extracting components from the 17.50 package. I guess it's time to do a bit of research :-).
Cheers,
Gary.
Quote:Gary Roberts wrote:
)
You seem pretty much on target Gary. I am completing about 4 to 5 CPU WUs just to get them out of the way. Then i will do 2 concurrent GPU WUs only.
robl wrote:... I am
)
I didn't notice much impact on GPU crunch times when running one or two CPU tasks as well. It does make a bit of difference to power draw though. With the older CPUs (2008 - 2010) that I've brought out of retirement I thought there would be an impact but there doesn't seem to be. I mostly run just one CPU task but on some quads I've been running two.
Cheers,
Gary.
You can run Einstein @ Home
)
You can run Einstein @ Home on Ubuntu 16.04 without having to install AMD drivers. The 4.10 kernel already comes with the amdgpu driver, and you just need to install the OpenCL packages: libclc-amdgcn, libclc-dev, libclc-r600, mesa-opencl-icd.
The problem is that you will not be able to run other projects like Milkyway (I haven't yet tested Seti), probably because it is OpenCL 1.1 and not 1.2 like the AMD drivers.
Short update: running E@H
)
Short update: running E@H full time on Ubu 16 and AMD RX 560 at a 2 GPU WUs concurrency now has me at about 230K avg credit per day. I did allow for a couple days of the new sky WUs to process but have stopped accepting any additional work for them.