Speaking of temperatures I checked the Pi’s that haven‘t had my fan mods and they were running between 69 and 72 degrees. That is with their top off and a full size desk fan blowing air onto their heatsink. I’ve set them to NNT for the time being. I need more 40mm fans and probably should get another case (or two) printed off.
Do you use a self-compiled app? Could you share your app or your compile-settings, app_info.xml etc
Yes, it's a self compiled app. And it requires some changes in the code. Some of them are really dirty hacks.
BOINC did'nt recognizes the Mali-GPU, so that I have to work around BOINC a little bit and didn't let BOINC know that I'm using the GPU.
Sharing the app with you: yes, but I need some time for testing and get everythink work properly. And my focus is more on getting the new GW-search running on my ARM-board's when it is released.
For the app-info there is a problem at the moment. I'm not able to tell BOINC to use a gpu-app and a cpu-app for BRP.
Not shure how to setup the app-info properly to allow both.
Do you use a self-compiled app? Could you share your app or your compile-settings, app_info.xml etc
Yes, it's a self compiled app. And it requires some changes in the code. Some of them are really dirty hacks.
BOINC did'nt recognizes the Mali-GPU, so that I have to work around BOINC a little bit and didn't let BOINC know that I'm using the GPU.
For a long long time, boinc here recognize the Mali GPUs (should I looking for it?), but there was not a BOINC Project around which supported this kind of GPU...
Sharing the app with you: yes, but I need some time for testing and get everythink work properly.
So my best wishes to you. I think many many users are waiting for a Mali-GPU-App...
BOINC did'nt recognizes the Mali-GPU, so that I have to work around BOINC a little bit and didn't let BOINC know that I'm using the GPU.
<snip>
Not shure how to setup the app-info properly to allow both.
You can define your own coprocessor in cc_config and then reference it in the app_info, however the project would need to do some config changes to send work to it. Given the small number of users with a Mali I expect it may not be worthwhile for the project.
The Mali T860 supports OpenCL 1.2 (and Vulcan 1.0) so the project could utilise their existing apps.
Finished in just 4333 sec on my Mali-T860 and only used 110sec of CPU-Time.
I am running your enhanced code for PI3s on several of mine. If you look at my "computers you can see the improvement over Pi2s. I am also running your Pi3 code on two Odroids. The Firefly looks interesting but not quite the $teel as a Pi3.
I was wondering were you had disappeared to. Now I know. Happy New Year to you and to all.
You can define your own coprocessor in cc_config and then reference it in the app_info, however the project would need to do some config changes to send work to it.
Thank you, I'ill give it a try.
PorkyPies wrote:
Given the small number of users with a Mali I expect it may not be worthwhile for the project.
Yes, this was only a playaround over the holidays for me.
PorkyPies wrote:
The Mali T860 supports OpenCL 1.2 (and Vulcan 1.0) so the project could utilise their existing apps.
Basiclly this is a recompile of the existing opencl-BRP-app. But I've changed some of the memory allocation to get use of the shared memory and get away with some of the copy's between the host and the device. There are still things in the opencl-kernels that can be optimized for the mali.
And yes that should work with all mali midgard(Txxx) and bifrost(Gxx) GPU's. But the the midgard gpus are using vector-operations and works best on vectors of 4. The BRP-opencl-code uses vector's of two. The T8xx handels this smaller vectors really well but older mali-gpu's like the T764 from the tinkerboard doesn't. So the performance of mali's < T8xx wouldn't be great. The new bifrost-gpu's works on scalar and vector-data so they would be optimal. But I only know one board that have a bifrost-gpu.
robl wrote:
I am running your enhanced code for PI3s on several of mine. If you look at my "computers you can see the improvement over Pi2s. I am also running your Pi3 code on two Odroids. The Firefly looks interesting but not quite the $teel as a Pi3.
Yes, it is pricey but they have these nice coreboards that cost "only" 95$ for 2gb and 115$ for the 4gb-model. My secret major goal is to design a "cluster-main-board" with a integrated GBE-switch were you can plug 16 of these board's in. Two GBE-Ports on the back of the mainboard that you can stack up multiple.
So I bought this board to get some performance number's to see if it's worth to design such a cluser-board. And yes this board is powerfull. The two A72-cores(@1.8Ghz) finishes a BRP-WU in around 6200sec. The A53-cores(@1.4Ghz) need's around 4.5h for a WU(not exactly testet yet). And I could use the GPU.
Powerfull yes but powerfull enough? Not shure now. Let me know what you are thinking?
robl wrote:
I was wondering were you had disappeared to. Now I know. Happy New Year to you and to all.
WhiteWulfe wrote:Out of sheer
)
I've checked the temps:
Pi ^4 = 54.2 degrees C with room temp of 29 degrees.
Modded Pi Case = 54.2 as well
MarksRpiCluster
Speaking of temperatures I
)
Speaking of temperatures I checked the Pi’s that haven‘t had my fan mods and they were running between 69 and 72 degrees. That is with their top off and a full size desk fan blowing air onto their heatsink. I’ve set them to NNT for the time being. I need more 40mm fans and probably should get another case (or two) printed off.
MarksRpiCluster
PorkyPies wrote:I've checked
)
Many thanks for the info! That's quite a difference between actively cooled and passive to say the least.
Just got my first validated
)
Just got my first validated result from a mobile GPU
https://einsteinathome.org/de/task/712910667
Finished in just 4333 sec on my Mali-T860 and only used 110sec of CPU-Time.
Hallo, could you tell us
)
Hallo,
could you tell us what for a ARM Board do you use?
Do you use a self-compiled app? Could you share your app or your compile-settings, app_info.xml etc?
Greetings
Peter
Peter Baetz schrieb:could you
)
It's a firefly RK3399 AIO
Yes, it's a self compiled app. And it requires some changes in the code. Some of them are really dirty hacks.
BOINC did'nt recognizes the Mali-GPU, so that I have to work around BOINC a little bit and didn't let BOINC know that I'm using the GPU.
Sharing the app with you: yes, but I need some time for testing and get everythink work properly. And my focus is more on getting the new GW-search running on my ARM-board's when it is released.
For the app-info there is a problem at the moment. I'm not able to tell BOINC to use a gpu-app and a cpu-app for BRP.
Not shure how to setup the app-info properly to allow both.
N30dG schrieb:Peter Baetz
)
N30dG wrote:BOINC
)
You can define your own coprocessor in cc_config and then reference it in the app_info, however the project would need to do some config changes to send work to it. Given the small number of users with a Mali I expect it may not be worthwhile for the project.
The Mali T860 supports OpenCL 1.2 (and Vulcan 1.0) so the project could utilise their existing apps.
MarksRpiCluster
N30dG wrote:Just got my first
)
I am running your enhanced code for PI3s on several of mine. If you look at my "computers you can see the improvement over Pi2s. I am also running your Pi3 code on two Odroids. The Firefly looks interesting but not quite the $teel as a Pi3.
I was wondering were you had disappeared to. Now I know. Happy New Year to you and to all.
PorkyPies schrieb:You can
)
Thank you, I'ill give it a try.
Yes, this was only a playaround over the holidays for me.
Basiclly this is a recompile of the existing opencl-BRP-app. But I've changed some of the memory allocation to get use of the shared memory and get away with some of the copy's between the host and the device. There are still things in the opencl-kernels that can be optimized for the mali.
And yes that should work with all mali midgard(Txxx) and bifrost(Gxx) GPU's. But the the midgard gpus are using vector-operations and works best on vectors of 4. The BRP-opencl-code uses vector's of two. The T8xx handels this smaller vectors really well but older mali-gpu's like the T764 from the tinkerboard doesn't. So the performance of mali's < T8xx wouldn't be great. The new bifrost-gpu's works on scalar and vector-data so they would be optimal. But I only know one board that have a bifrost-gpu.
Yes, it is pricey but they have these nice coreboards that cost "only" 95$ for 2gb and 115$ for the 4gb-model. My secret major goal is to design a "cluster-main-board" with a integrated GBE-switch were you can plug 16 of these board's in. Two GBE-Ports on the back of the mainboard that you can stack up multiple.
So I bought this board to get some performance number's to see if it's worth to design such a cluser-board. And yes this board is powerfull. The two A72-cores(@1.8Ghz) finishes a BRP-WU in around 6200sec. The A53-cores(@1.4Ghz) need's around 4.5h for a WU(not exactly testet yet). And I could use the GPU.
Powerfull yes but powerfull enough? Not shure now. Let me know what you are thinking?
Was here all the time ;).
Happy new year to you and all other.