The Nvidia software stack is definitely not open. It just works well.Tegra might be open, but it's a power hungry beast.
Another difficulty is releasing open-source GPU drivers attracts patent trolls.If the IP doesn't belong to you then you can't release it (unless you like ending up in court or paying large penalties).
In the race to the bottom I give credit to ATI for helping with the open source Radeon drivers for the R100 class cards twenty years ago. The source was closed in connection with the purchase of ATI by AMD and reopened a number of years later when AMD realized how important that was for Linux.
My impression is one of the reasons there are AMD Radeon GPUs in the Frontier supercomputer at Oak Ridge National Laboratory is due to the open-source nature of the ROCm software stack and Linux drivers. It's notable that El Capitan and LUMI also use AMD accelerators. I think the Linux driver for the Intel GPU Max in the Aurora supercomputer is open source but am not sure about all the libraries.
Of course many also employ the closed-source Nvidia H100 accelerators. While I haven't made the pie chart, my guess these days is more scientific computing power comes from supercomputers without Nvidia hardware than with. Training AI is a different story.
From an educational point of view, it would be nice if the Pi 5 provided a software stack to learn similar kinds of heterogeneous GPU computation as common on x86 systems.
Statistics: Posted by ejolson — Fri Jan 26, 2024 9:54 pm