ISAR evaluation nVidia image generator based on Debian 11 (bullseye)
Dear all,
I am proud to announce the v0.9.2 of an ISAR evaluation nVidia image
generator based on Debian bullseye. I hardly believe that this project
will surprise you because of the technology adopted nor its use but
probably you will find this file interesting, in particular the
rationale chapter (added inline at the bottom for sake of completeness).
https://github.com/robang74/isar-nvidia-debian/blob/main/README.md#rationale
It is an original rework of code based only on published source code
and recipes. At the moment, this project is locked on a specific
version of the driver (515.65.01) and CUDA libraries (11.7). It is my
desire that it remains in this state for a purpose because it is just
a proof of concept (PoC) and not even a commercial demo. Thanks in
advance.
I have no hardware with a supported nVidia card. It is appreciated
that someone will take a nvidia-smi test on a complete or nvdocker
image. The feedback would be appreciated also in private form. For
this reason this project should be considered untested.
After all, at the console login screen and immediately after a
message of no warranty granted but delivered AS-IS is always
displayed.
Have fun! <3
Best regards, R-
====
Rationale
---------
An equivalent result can be obtained installing a Debian 11, adding the nVidia
repositories dedicated to the developers and the other one dedicated to the
docker runtime, then installing the 'cuda-demo-suite-11-7' and 'nvidia-docker2'
packages.
The most sensitive difference between these two approaches is that the ISAR
image contains the open-source driver while the apt installed the closed-source.
In fact, this project is a proof-of-concept that shows how to add the
open-source
nVidia driver in a Debian 11 system integrating it with the proprietary full
software stack without violating the licence and being able to redistribute
the image, at least for some usages allowed by the licences (\*).
- https://opensource.stackexchange.com/questions/10082/geforce-nvidia-driver-license-for-commerical-use
This project aims to provide a way to deliver a system with nVidia full stack
software installed which is legally distributable also for commercial uses.
- https://www.nvidia.com/en-us/drivers/unix
In fact, up today (515.76) the .run archive that contains the driver and the
CUDA libraries is licenced in a way for which two essential operations are not
permitted:
- §2.1.2 does not allow the compilation essential for deliver a binary driver
- §2.1.3 does not allow to repackage the .run content in many .deb packages
This project works around these limitations using the open-source driver
- https://github.com/NVIDIA/open-gpu-kernel-modules
in order to not violate the §2.1.2 and installing the nVidia software from their
public repositories without changing the .deb packages content and removing just
few dependencies - which are just text fields into a .deb architecture and have
nothing to do with the content delivered aka package metadata, only - allows to
avoid installing the closed-source driver and the related packages.
This allows also to choose a complete different kernel version respect the
one delivered with the Debian 11 and compile it by an ISAR recipe applying
a custom configuration and patches like this one:
- https://lore.kernel.org/lkml/20220921063638.2489-1-kprateek.nayak@amd.com
that unlock AMD Ryzen CPUs a more +51% of computation power lost due
to an old bug.
(\*) **Legal notes**
- no any warranty is granted and further license changes might happen.
- debian legal ml https://lists.debian.org/debian-legal/2022/10/msg00004.html
Reply to: