WCG/BOINC GPU.

Keith,

  I'm not aware of BOINC clients available for ARM.  The one I know of is for MS WIndows-64.

The RPi GPU doesn't currently support GPGPU/OpenCL, so is difficult to use for general-purpose computing.

The RPi CPU is about the speed of a Pentium 2 300MHz.

The GPU uses up to 128MB of memory, all of which is shared with the CPU.  Actually the CPU

accesses memory through the GPU, including accessing the L2 cache.

There is a partial datasheet from Broadcom available, and also some information

on a recent RPi forum thread about getting X-Window drivers to work with the GPU.

http://www.element14.com/community/docs/DOC-43016/l/broadcom-datasheet-for-bcm2835-soc-used-in-raspberry-pi

http://www.raspberrypi.org/forum/general-discussion/so-whos-workingworked-on-an-xorg-server

Keith,

  I'm not aware of BOINC clients available for ARM.  The one I know of is for MS WIndows-64.

The RPi GPU doesn't currently support GPGPU/OpenCL, so is difficult to use for general-purpose computing.

The RPi CPU is about the speed of a Pentium 2 300MHz.

The GPU uses up to 128MB of memory, all of which is shared with the CPU.  Actually the CPU

accesses memory through the GPU, including accessing the L2 cache.

There is a partial datasheet from Broadcom available, and also some information

on a recent RPi forum thread about getting X-Window drivers to work with the GPU.

http://www.element14.com/community/docs/DOC-43016/l/broadcom-datasheet-for-bcm2835-soc-used-in-raspberry-pi

http://www.raspberrypi.org/forum/general-discussion/so-whos-workingworked-on-an-xorg-server

BOINC client is open source and looks easy enough to port/recompile to ARM. Its made for Linux and in fact its the Windows version that has to be cross compiled.

As you say the GPU does not support GPGPU/OpenCL. Seems most projects that support WCG are using CUDU which is specific to NVIDIA.

Guess will have to wait to see if Broadcom decide to support OpenCL at some time in the future.

Pity I would have happily coughed up £30.00 to have a Raspberry Pi on my home network doing grid computing work and  only using minimum electricity.

I agree with Eben Upton on this, that the Raspberry Pi is not likely to be an effective board for HPC applications.  Its CPU is slow, there is no OpenCL for its GPU, and the GPU is closed so that nobody but Broadcom could write a generic tasking library for it.  I guess reverse engineering the GPU is still possible, but nobody has started on such a project yet, as far as we know.

That said, there is one possible way to use Raspberry Pi in HPC, but it requires some engineering knowledge.  Interface an FPGA directly to the Rpi hardware, and program up the innermost loops of your HPC algorithms in VHDL or Verilog.  As long as your computing time is much larger than your communication time over all the comms interfaces, your application will still be able to run efficiently.  (Power efficiency will be lost though.)

We had a thread running about Rpi + FPGA a little while ago -- http://www.element14.com/community/thread/17692?tstart=0 , and also one about Rpi clusters which is slightly related -- http://www.element14.com/community/thread/17542?tstart=0  .

Morgaine.