CUDA 2.1 on unsupported distros

To whom it may concern: Things seem to runs smoothly (my regression apps, not the SDK as these are headless nodes and I can’t run the examples that require a display) on unsupported platforms as long as the system GCC of a supported platform can be sneaked in somehow.

In detail: I’ve been running all these tests on an unsupported suse 10.2 platform. I compiled gcc 4.3.1 (a specialisation of which ships with [supported] OpenSUSE 11.0) and 4.3.2 [not part of suse 11.0] myself and got all paths right through the magic of the module environment. I combine these gccs with the downloadable toolkit for suse 11.0 and Intel icc 10.1 for the Fortran bit of my application. My non-Fortran regression tests that capture more than what my application needs run fine. My Fortran applications run fine. On 32 and 64 bit linux. Yehaw.

It makes an awful lot of sense for you guys at NVIDIA to release CUDA on a per-distro basis. My suggestion is to additionally release a fallback version that is officially confirmed to work with a given gcc trunk. As in “for CUDA 2.1, use the display driver provided, compile your own gcc x.y.z, and the SDK apps will work. Google how to compile your own gcc”.