Thanks Paul, On 2024-04-22 02:19, Paul Wise wrote:
I created that page, but no-one has ever added any wishlist items to it, and it doesn't get many mentions on Debian lists except from me, so I am not sure how useful the page is or how many people see it. It might be more useful to put a callout for donations on d-d-a via DevNews, or on micronews.d.o or other publicity team resources. https://wiki.debian.org/DeveloperNews https://wiki.debian.org/Teams/Publicity/micronews https://wiki.debian.org/Teams/Publicity
That's a good idea. I've had success with reaching out to component manufacturers directly, but it's a rather time-consuming process and there are lots of expenses that cannot be reasonably covered in that way.
There are many people interested in having this sort of test coverage. As far as I know, there is no other organization that is publishing test results for the AMD GPU compute stack on as many different hardware architectures as Debian is. I hope that will inspire people to contribute to support the effort.
Working with some rough numbers, I think we need roughly 5k USD to add test coverage of gfx900, gfx90c, gfx1036, and gfx1103 APUs. We have the actual APUs themselves already, but we need four servers worth of components and accessories.
Over the past year, I've personally spent around 14k USD on building the Debian ROCm CI, of which around 5k USD has been reimbursed out of general Debian project funds. I believe Christian has had a few thousand dollars in costs as well. The Debian project funds have been of great help in kickstarting the system, but I do not want the ROCm CI to be a financial burden. It should pull its own weight, inspiring donations to Debian roughly equal to its costs.
We have partnerships with hardware companies and universities that cover many of the most expensive aspects of building and running our servers. My original budget plan was approximately 25k USD per year, but it now seems feasible to build and run the system with 10k USD per year.
We're obviously not going to buy an MI300 system with that sort of budget, so we would still rely heavily on hardware donations. However, we would be able afford to purchase supporting accessories (e.g., a used switch and a few network transceivers), replace any failed parts in our existing servers, and buy parts for new systems to expand hardware coverage.
Sincerely, Cory Bloor