Is there a "turnkey" ARM computer system (besides a Tablet) that you would recommend.

I have now run across two ARM development boxes. One with 24 cores@$1200 and One that doesn't have a production date but is "supposed" to be $600-800.

Tom

---

A proud member of the OFA (Old Farts Association).

Development systems are not always useable as they stand, but require some external support - so be very careful when deciding which way to jump.

For a starter the Raspberry Pi family are far more "complete" than many others - they come with operating system, HMI interfaces (mouse, keyboard, video) - the only things needed are a PSU and a means of getting data in and out (I use a cheap keyboard and mouse, plunging a monitor in when needed). There are tools to allow several Pi to be used as a single computer - I don't know the limits, but given each Pi-3 B has for cores the core count grows quite rapidly and for a very low cost. Admittedly a single core is not a road runner, but they do work out of the box.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Is there a "turnkey" ARM computer system (besides a Tablet) that you would recommend.

I have now run across two ARM development boxes. One with 24 cores@$1200 and One that doesn't have a production date but is "supposed" to be $600-800.

Tom

Well, the fastest ARM processor that there is a Seti app for is the ARMv8 processor for SETI@home v8 v8.02 aarch64-unknown-linux-gnu

That processor is faster than the ARMv7 processors in the Raspberry Pi cohort. You can compare my RPi3 and Nano hosts for cpu crunching times.

I just searched on ARMv8 computers and landed on the Gigabyte ThunderXstation and the Cavium ThunderX built with the ThunderX2 ARMv8 processor.
This is a pdf on the processor https://origin-www.marvell.com/documents/cmvd78bk8mesogdusz6t/

2.2GHz 32-core 4-threads per core and support for lots of PCIe lanes. Interesting.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

2.2GHz 32-core 4-threads per core and support for lots of PCIe lanes. Interesting.

If fast crunching is you aim, go with a GPU of some kind.

If low cost low power crunching is your aim, then ARM it is.

Oh, IIRC GCC itself doesn't have all the optimizations that it has on AMD/Intel. Its still a work in progress.

Of course YMMV and new stuff comes out all the time.

If either myself or Sam is ever successful in compiling gpu apps on the aarch64-unknown-linux-gnu platform, then that motherboard and processor would make a great platform to drive lots of gpus.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Just a quick FYI. Both CyborgSam and myself have compiled the zi3v special app to run on the Nvidia Jetson Nano SBC with its integrated Tegra X1 gpu. Most tasks run in 30 - 60 minutes compared to the stock cpu app which takes 4 - 8 hours per task.
Here is an example. https://setiathome.berkeley.edu/result.php?resultid=7765593986

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

You're going to have to change over to using the "anonymous platform" as Keith does to get them, though I'm sure that Keith will supply all the info on doing that. ;-)

Cheers.

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

I'll have to investigate the VC4CL library, Hadn't heard of it.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

Well, my Raspberry 3B+ apparently supports OpenCL now so I'd like to find a way to get some work for its GPU. I guess the stock seti arm application only uses the CPU.

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

Well, my Raspberry 3B+ apparently supports OpenCL now so I'd like to find a way to get some work for its GPU. I guess the stock seti arm application only uses the CPU.

Can you post the first startup lines of BOINC in the Event Log showing the OpenCL driver that BOINC detects.

You will have to get the sah_v7_opt and Xbranch directories from the BOINC SVN repository. Then install all the ARM development libraries and then run through the autosetup and configure procedures and then hope that you are successful with compiling the SoG app. Expect to spend a lot of time debugging the code and turning all kinds of things off and make changes. The source code is intended for x86_64, not ARM.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

It might be possible to compile the stock cpu app to be run on the gpu. That is what I have done for the Einstein BRP4 cpu application on my Nano. But that was to make the Nano Maxwell gpu run the cpu code as a CUDA application.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

Well, my Raspberry 3B+ apparently supports OpenCL now so I'd like to find a way to get some work for its GPU. I guess the stock seti arm application only uses the CPU.

Can you post the first startup lines of BOINC in the Event Log showing the OpenCL driver that BOINC detects.

You will have to get the sah_v7_opt and Xbranch directories from the BOINC SVN repository. Then install all the ARM development libraries and then run through the autosetup and configure procedures and then hope that you are successful with compiling the SoG app. Expect to spend a lot of time debugging the code and turning all kinds of things off and make changes. The source code is intended for x86_64, not ARM.

This is what BOINC detects:
OpenCL: VideoCore IV GPU 0: VideoCore IV GPU (driver version 0.4.9999, device version OpenCL 1.2 VC4CL 0.4.9999, 128MB, 128MB available, 0 GFLOPS peak)

I tried to compile the optimized branch you mentioned, and the result was every error known to mankind. It's difficult to fix code you aren't familiar with. I don't know if they have prepared the code for ARM at all, there is a configure commandline for ARM included, but it doesn't work very well on my Raspberry at least.

It might be possible to compile the stock cpu app to be run on the gpu. That is what I have done for the Einstein BRP4 cpu application on my Nano. But that was to make the Nano Maxwell gpu run the cpu code as a CUDA application.

This sounds very interesting. Perhaps it would be easier to go that route than trying to compile the optimized branch?

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

Well, my Raspberry 3B+ apparently supports OpenCL now so I'd like to find a way to get some work for its GPU. I guess the stock seti arm application only uses the CPU.

Can you post the first startup lines of BOINC in the Event Log showing the OpenCL driver that BOINC detects.

You will have to get the sah_v7_opt and Xbranch directories from the BOINC SVN repository. Then install all the ARM development libraries and then run through the autosetup and configure procedures and then hope that you are successful with compiling the SoG app. Expect to spend a lot of time debugging the code and turning all kinds of things off and make changes. The source code is intended for x86_64, not ARM.

This is what BOINC detects:
OpenCL: VideoCore IV GPU 0: VideoCore IV GPU (driver version 0.4.9999, device version OpenCL 1.2 VC4CL 0.4.9999, 128MB, 128MB available, 0 GFLOPS peak)

I tried to compile the optimized branch you mentioned, and the result was every error known to mankind. It's difficult to fix code you aren't familiar with. I don't know if they have prepared the code for ARM at all, there is a configure commandline for ARM included, but it doesn't work very well on my Raspberry at least.

Ha ha LOL, Really, sorry about that.

Yes it is a bear to compile the x86 code for ARM. Had to throw out basically everything for configure that was provided in the source code and come up with our own configures. I think Sam and I have spent over a month on it so far. Might not be so bad for ARMv7a of the RPi compared to our ARMv8 of the Nano since that is a long running mainstream ARM processor.

I also have serious doubt that you can get any OpenCL gpu app to run in 128 MB of GPU RAM

I have 4 GB of shared memory in my Nano.

[Edit] This is what my Nano CUDA tasks see for the application

setiathome_CUDA: Found 1 CUDA device(s):
  Device 1: NVIDIA Tegra X1, 3956 MiB, regsPerBlock 32768
     computeCap 5.3, multiProcs 1

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

I was wondering if anyone ever had any luck in compiling the optimized branch with OpenCL for ARM? I installed VC4CL (OpenCL GPU support) on my Raspberry 3B+ and now Boinc recognizes the VideoCore GPU, so it would be interesting to get some work for it.

We looked and Nvidia does not support any OpenCL on the Nano.

Well, my Raspberry 3B+ apparently supports OpenCL now so I'd like to find a way to get some work for its GPU. I guess the stock seti arm application only uses the CPU.

Can you post the first startup lines of BOINC in the Event Log showing the OpenCL driver that BOINC detects.

You will have to get the sah_v7_opt and Xbranch directories from the BOINC SVN repository. Then install all the ARM development libraries and then run through the autosetup and configure procedures and then hope that you are successful with compiling the SoG app. Expect to spend a lot of time debugging the code and turning all kinds of things off and make changes. The source code is intended for x86_64, not ARM.

This is what BOINC detects:
OpenCL: VideoCore IV GPU 0: VideoCore IV GPU (driver version 0.4.9999, device version OpenCL 1.2 VC4CL 0.4.9999, 128MB, 128MB available, 0 GFLOPS peak)

I tried to compile the optimized branch you mentioned, and the result was every error known to mankind. It's difficult to fix code you aren't familiar with. I don't know if they have prepared the code for ARM at all, there is a configure commandline for ARM included, but it doesn't work very well on my Raspberry at least.

Ha ha LOL, Really, sorry about that.

Yes it is a bear to compile the x86 code for ARM. Had to throw out basically everything for configure that was provided in the source code and come up with our own configures. I think Sam and I have spent over a month on it so far. Might not be so bad for ARMv7a of the RPi compared to our ARMv8 of the Nano since that is a long running mainstream ARM processor.

I also have serious doubt that you can get any OpenCL gpu app to run in 128 MB of GPU RAM

I have 4 GB of shared memory in my Nano.

[Edit] This is what my Nano CUDA tasks see for the application

setiathome_CUDA: Found 1 CUDA device(s):
  Device 1: NVIDIA Tegra X1, 3956 MiB, regsPerBlock 32768
     computeCap 5.3, multiProcs 1

Hehe 256MB is the limit in raspi-config, probably possible to go higher in config.txt? But the CPU app doesn't need much memory, I run seti on a Raspberry pi Zero and it has only 512mb RAM. Would a GPU app crave more memory than the CPU app?

Hehe 256MB is the limit in raspi-config, probably possible to go higher in config.txt? But the CPU app doesn't need much memory, I run seti on a Raspberry pi Zero and it has only 512mb RAM. Would a GPU app crave more memory than the CPU app?

Absolutely, the whole point of computing on gpus is running parallel computations. That takes memory for the temp compute results in the arrays. I think the smallest amount of gpu ram that any of the gpu apps can run on is 2GB. Or at least I think I've seen posts to that fact. Willing to be corrected otherwise.

You can run a swap file on external hard drive on the Pi. That might make it possible.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)

Absolutely, the whole point of computing on gpus is running parallel computations. That takes memory for the temp compute results in the arrays. I think the smallest amount of gpu ram that any of the gpu apps can run on is 2GB. Or at least I think I've seen posts to that fact. Willing to be corrected otherwise.

You can run a swap file on external hard drive on the Pi. That might make it possible.

The 3B+ has some limitations with only 1GB RAM of course, I'll use a swap file if needed. I'll deal with memory management later :-) For now I'd be happy just to be able to compile something that will run. If it's possible to adapt the stock cpu app to make it run on the GPU then I can work with that. You mentioned you had done something similar with the Einstein cpu application?

Yes there isn't any normal BRP4G tasks anymore, for the app that was gpu based at Einstein. But they are still creating the cpu tasks for BRP4. So I have an app that crunches those on the Nano gpu. It is CUDA based.

Basically you take the cpu source code and setup your compiling environment and make the target the gpu instead of the cpu.

---

Seti@Home classic workunits:20,676 CPU time:74,226 hours

A proud member of the OFA (Old Farts Association)