I'm interested in playing with this, but I'm not interested in paying $3000
for a high-end nVidia NPU. It seems to be in principle possible to split the task into smaller subtasks that could be distributed to a Beowulf-style Pi cluster. Several enclosures exist that let you connect multiple Compute
Modules over a high-speed bus, I hear.

I have minimal experience with Raspberry Pi (not zero, but minimal). I have none with setting up Beowulf clusters, and none with decomposing machine learning tasks and distributing them among processors. Thus, I wonder if
there might be an existing project I could learn from and maybe even
eventually contribute to, even if only as a tester.

Thanks.
--
Carl Fink carl@finknetwork.com https://reasonablyliterate.com https://nitpicking.com If you want to make a point, somebody will take the point and stab you with it.
-Kenne Estes

--- SoupGate-Win32 v1.05
* Origin: Agency HUB, Dunedin - New Zealand | Fido<>Usenet Gateway (3:770/3)

On 12/24/24 09:35, Carl Fink wrote:

I'm interested in playing with this, but I'm not interested in paying $3000 for a high-end nVidia NPU. It seems to be in principle possible to split the task into smaller subtasks that could be distributed to a Beowulf-style Pi cluster. Several enclosures exist that let you connect multiple Compute Modules over a high-speed bus, I hear.

I have minimal experience with Raspberry Pi (not zero, but minimal). I have none with setting up Beowulf clusters, and none with decomposing machine learning tasks and distributing them among processors. Thus, I wonder if there might be an existing project I could learn from and maybe even eventually contribute to, even if only as a tester.

Thanks.

There were some projects that used multiple PIs. From a cost and
complexity standpoint they were more "because we can" rather than
practical. A multicore AMD or INTEL processor would be a better option.
If you skip the high power GPU the system cost is lower. Also the MPI
cluster code is "off the shelf" for those processors.

https://mpitutorial.com/tutorials/mpi-hello-world/

--- SoupGate-Win32 v1.05
* Origin: Agency HUB, Dunedin - New Zealand | Fido<>Usenet Gateway (3:770/3)

On 2024-12-24, Dennis <dennis@none.none> wrote:

There were some projects that used multiple PIs. From a cost and
complexity standpoint they were more "because we can" rather than
practical. A multicore AMD or INTEL processor would be a better option.
If you skip the high power GPU the system cost is lower. Also the MPI
cluster code is "off the shelf" for those processors.

https://mpitutorial.com/tutorials/mpi-hello-world/

Thank you.
--
Carl Fink carl@finknetwork.com https://reasonablyliterate.com https://nitpicking.com If you want to make a point, somebody will take the point and stab you with it.
-Kenne Estes

--- SoupGate-Win32 v1.05
* Origin: Agency HUB, Dunedin - New Zealand | Fido<>Usenet Gateway (3:770/3)