I have moved the Fury X to another host with PCIe 3.0.
BRP6 times are reduced now from 2700 to 2500 s, but this is still quite behind your Nano's ~2200 s.
I'll try to suspend all other CPU tasks to see if/what impact can that have...
Note, that the new host https://einsteinathome.org/host/7192129/tasks has also a HD 7950 (not 2xFiji as reported by BOINC), so the slower times come from this GPU.
EDIT: Running without other CPU tasks gives ~2300 s. Still not what I expected..
I guess the only thing left to blame is the OS - I'm running Win10.
Howdy Mumak,
The biggest impact on my Nano was changing the CPU speed
which surprised me. I didn't think it would matter (that much)
I overclocked the CPU from 3.3 to 3.95ghz and the (Parkes PMPS XT)
units dropped from 42 min to about 36-37 minutes to complete.
Last night in an effort to run cooler I dropped the power
level -10% (in the crimson settings) now, it's running
63- 64 Celsius and, crunching in about 38 min. (2280)
So, surprisingly, the biggest difference was the CPU speed.
I tried adjusting main memory speed (with little to Zero effect)
Now, I'm experimenting with running a Grav Wave search O1 all-sky tuning
unit or two simultaneously.
Bill
-edit- previously, all the CPU was doing was running the GPU )
Yup, CPU performance has a huge impact on these OpenCL tasks - you can see that they spend >80% of time in CPU cycles... CUDA tasks run with much less CPU time (<10%).
So in the end we get quite comparable numbers (if we consider our CPU settings). You get higher temperatures, but lower power (and slightly reduced clock because of that) and I'm at full power, but lower temp (<= 50 C).
I don't know how much power exactly does the GPU consume - I can access and read out the GPU VRs, but these Fiji's are problematic and such access has a high probability of causing a total system crash.
Howdy Mumak,
The biggest impact on my Nano was changing the CPU speed
which surprised me. I didn't think it would matter (that much)
I overclocked the CPU from 3.3 to 3.95ghz and the (Parkes PMPS XT)
units dropped from 42 min to about 36-37 minutes to complete.
Last night in an effort to run cooler I dropped the power
level -10% (in the crimson settings) now, it's running
63- 64 Celsius and, crunching in about 38 min. (2280)
So, surprisingly, the biggest difference was the CPU speed.
I tried adjusting main memory speed (with little to Zero effect)
Now, I'm experimenting with running a Grav Wave search O1 all-sky tuning
unit or two simultaneously.
Bill
-edit- previously, all the CPU was doing was running the GPU )
What driver are you using? My first run on my R9 390x using Cat 15.7 had run times ~36 min like your R9 Nano for Parkes tasks. However after changing to 15.12 I'm seeing ~45 min. It could just be the different batch of work. I don't know how much variance these tasks have yet.
Also I should add my i5-4670K is running no CPU tasks and is at 4.0GHz.
What driver are you using? My first run on my R9 390x using Cat 15.7 had run times ~36 min like your R9 Nano for Parkes tasks. However after changing to 15.12 I'm seeing ~45 min. It could just be the different batch of work. I don't know how much variance these tasks have yet.
Also I should add my i5-4670K is running no CPU tasks and is at 4.0GHz.
Howdy Hal,
I'm running Radeon Crimson 15.11.1 beta right now ..Why ? .. too lazy to download and install a newer one :-)
You are absolutely right - the driver can make a big difference.
I am also reluctant to try the Vulkan driver - I'm keeping an eye
on that over at Seti to see how they fare )
I have upgraded to 16.2.1 and all is well.
I still think that the usage is fluctuating too much and too much CPU resources used. Hope AMD will fix that in later drivers...
I have upgraded to 16.2.1 and all is well.
I still think that the usage is fluctuating too much and too much CPU resources used. Hope AMD will fix that in later drivers...
I hear ya !
I see you didn't want to gamble on those AMD Vulkan drivers either )
I don't know if HAL9000 is testing those but, some are over at Seti :)
The performance drop is
)
The performance drop is probably because the mainboard supports PCIe 2.0 only.
-----
I have moved the Fury X to
)
I have moved the Fury X to another host with PCIe 3.0.
BRP6 times are reduced now from 2700 to 2500 s, but this is still quite behind your Nano's ~2200 s.
I'll try to suspend all other CPU tasks to see if/what impact can that have...
Note, that the new host https://einsteinathome.org/host/7192129/tasks has also a HD 7950 (not 2xFiji as reported by BOINC), so the slower times come from this GPU.
EDIT: Running without other CPU tasks gives ~2300 s. Still not what I expected..
I guess the only thing left to blame is the OS - I'm running Win10.
-----
Howdy Mumak, The biggest
)
Howdy Mumak,
The biggest impact on my Nano was changing the CPU speed
which surprised me. I didn't think it would matter (that much)
I overclocked the CPU from 3.3 to 3.95ghz and the (Parkes PMPS XT)
units dropped from 42 min to about 36-37 minutes to complete.
Last night in an effort to run cooler I dropped the power
level -10% (in the crimson settings) now, it's running
63- 64 Celsius and, crunching in about 38 min. (2280)
So, surprisingly, the biggest difference was the CPU speed.
I tried adjusting main memory speed (with little to Zero effect)
Now, I'm experimenting with running a Grav Wave search O1 all-sky tuning
unit or two simultaneously.
Bill
-edit- previously, all the CPU was doing was running the GPU )
Yup, CPU performance has a
)
Yup, CPU performance has a huge impact on these OpenCL tasks - you can see that they spend >80% of time in CPU cycles... CUDA tasks run with much less CPU time (<10%).
So in the end we get quite comparable numbers (if we consider our CPU settings). You get higher temperatures, but lower power (and slightly reduced clock because of that) and I'm at full power, but lower temp (<= 50 C).
I don't know how much power exactly does the GPU consume - I can access and read out the GPU VRs, but these Fiji's are problematic and such access has a high probability of causing a total system crash.
-----
Right now I'm running 1 GPU
)
Right now I'm running 1 GPU task and 1 CPU task simultaneously.
Total system power draw is 283 watts (+ - 3 watts)
It would drop 10 watts by shutting down the CPU task.
Bill
-edit- My ancient Core 2 machine w HD 7970
draws about 285 watts - (GPU tasks only)
RE: Howdy Mumak, The
)
What driver are you using? My first run on my R9 390x using Cat 15.7 had run times ~36 min like your R9 Nano for Parkes tasks. However after changing to 15.12 I'm seeing ~45 min. It could just be the different batch of work. I don't know how much variance these tasks have yet.
Also I should add my i5-4670K is running no CPU tasks and is at 4.0GHz.
RE: What driver are you
)
Howdy Hal,
I'm running Radeon Crimson 15.11.1 beta right now ..Why ? .. too lazy to download and install a newer one :-)
You are absolutely right - the driver can make a big difference.
I am also reluctant to try the Vulkan driver - I'm keeping an eye
on that over at Seti to see how they fare )
Bill
I have upgraded to 16.2.1 and
)
I have upgraded to 16.2.1 and all is well.
I still think that the usage is fluctuating too much and too much CPU resources used. Hope AMD will fix that in later drivers...
-----
RE: I have upgraded to
)
I hear ya !
I see you didn't want to gamble on those AMD Vulkan drivers either )
I don't know if HAL9000 is testing those but, some are over at Seti :)
Bill
There's no reason why the
)
There's no reason why the Vulkan drivers should perform better in OpenCL.
-----