My 570 is taking less than 20 minutes to run Two tasks at a time, the NV 2070s take about 10.5 minutes to run One task at a time, or, over 20 minutes for two at a time.
The R7 is a beast. Just over 3mins per task is great. Einstein’s current OpenCL apps definitely give an edge to AMD cards. Not sure if the apps here are more sensitive to memory bandwidth or DP performance, hard to say since the R7 excels in both of those areas.
My 570 is taking less than 20 minutes to run Two tasks at a time, the NV 2070s take about 10.5 minutes to run One task at a time, or, over 20 minutes for two at a time.
I just got two Rx 570's online and and they are munching around 9.5 minutes per Pulsar#1 task (I think).
. . Less than 10 mins per task is impressive. My GTX1060-6GB cards are taking just under 19 mins so are only half as fast with the OpenCl app. But AMD have been quicker than Nvidia under OpenCl for quite a while. But I have NV cards in all 5 rigs yet they are running different versions of the app. Does anyone know why E@H picks the nvidia-1k version of V1.20 for some cards and V1.22 for others?
Until they created the nvidia-TV application, all work bombed out on Turing-Volta cards, hence the two applications available, the older nvidia-1K for Pascal and earlier and the TV for Turing and newer.
. . But that is what surprises me. All my cards are Pascal or older yet the 1050 is running V1.22 while the others are all running the V1.20-1K version. Yet it only has 2Gb ram while most of the others are 4GB or better. I am puzzled.
There might be other selection criteria that the scheduler uses other than CC rating. That is what was bombing out the nvidia-1k application when the Volta and Turing cards came out with their CC = 7.5 rating. It didn't know how to handle that new value.
Could be the limited memory is the next criteria that the scheduler uses. No way of knowing since the scientists never post here about how things work. We can can only guess by observation.
Can someone explain the
)
Can someone explain the highly variable results on the Radeon VII gpu(s)?
In the Leader board list I see Pulsar#1 tasks running as low as 6 minutes and as high as 20 minutes.
I am also getting 20 minute results from my Rx 560's when I run them with 0.9 cpu and 1 gpu.
So am I seeing that much difference between Linux/Windows? Or that much difference between air cooled and water cooled?
Thanks.
Tom M
A Proud member of the O.F.A. (Old Farts Association).
Some users may use some sort
)
Some users may be running their GPUs under some sort of power limiting. I's difficult to compare run times unless conditions are known.
Richie wrote: Some users may
)
also running multiple tasks per card. Making runtimes appear longer than baseline.
id day if you see one card running 6min/task and another card doing 20/task, the latter is probably doing 3-4x tasks per GPU.
_________________________________________________________________________
This one is running One task
)
This one is running One task at a time, it should give you a clue about the others that are showing less than twice as long, https://einsteinathome.org/host/12780933/tasks/0/0?page=273
My 570 is taking less than 20 minutes to run Two tasks at a time, the NV 2070s take about 10.5 minutes to run One task at a time, or, over 20 minutes for two at a time.
The R7 is a beast. Just over
)
The R7 is a beast. Just over 3mins per task is great. Einstein’s current OpenCL apps definitely give an edge to AMD cards. Not sure if the apps here are more sensitive to memory bandwidth or DP performance, hard to say since the R7 excels in both of those areas.
_________________________________________________________________________
TBar wrote:This one is
)
I just got two Rx 570's online and and they are munching around 9.5 minutes per Pulsar#1 task (I think).
https://einsteinathome.org/host/12830231
Tom M
A Proud member of the O.F.A. (Old Farts Association).
Hi Tom . . Less than
)
Hi Tom
. . Less than 10 mins per task is impressive. My GTX1060-6GB cards are taking just under 19 mins so are only half as fast with the OpenCl app. But AMD have been quicker than Nvidia under OpenCl for quite a while. But I have NV cards in all 5 rigs yet they are running different versions of the app. Does anyone know why E@H picks the nvidia-1k version of V1.20 for some cards and V1.22 for others?
Stephen
Until they created the
)
Until they created the nvidia-TV application, all work bombed out on Turing-Volta cards, hence the two applications available, the older nvidia-1K for Pascal and earlier and the TV for Turing and newer.
. . Thanks Keith, .
)
. . Thanks Keith,
. . But that is what surprises me. All my cards are Pascal or older yet the 1050 is running V1.22 while the others are all running the V1.20-1K version. Yet it only has 2Gb ram while most of the others are 4GB or better. I am puzzled.
Stephen
? ?
There might be other
)
There might be other selection criteria that the scheduler uses other than CC rating. That is what was bombing out the nvidia-1k application when the Volta and Turing cards came out with their CC = 7.5 rating. It didn't know how to handle that new value.
Could be the limited memory is the next criteria that the scheduler uses. No way of knowing since the scientists never post here about how things work. We can can only guess by observation.