[Bf-committers] How much ram is needed for compiling cuda kernels ?
troubledaemon at gmail.com
Fri Apr 19 14:49:09 CEST 2013
I use ccmake (the ncurses version of cmake-menu) here in Linux. For my
cuda compiling, I generally leave WITH_CYCLES_CUDA_BINARIES set to
OFF, which has the effect of delaying building of the cuda kernel
until I actually run blender and enable cycles for the first time.
After the first run, this should have the effect of only comping the
cuda that is actually needed for your system.
If you would rather build your kernel during compile instead, set that
above line to ON, and then in advanced settings, change
CYCLES_CUDA_BINARIES_ARCH to sm_21 only (the newer default is:
"sm_20;sm_21;sm_30" for me atm, but I think it had sm_13 in there
Hope this helps, gl o/
On Fri, Apr 19, 2013 at 8:31 AM, Marko Radojcic <sambucuself at gmail.com> wrote:
> Thanks. This is the answer I was looking for.
> My GT 430 has 2.1 compute capability so 1.3 is not needed for me, and
> apparently I didn't force the sm_21 value properly.
> What shoud I edit to force compiling only 2.1 cuda kernel ?
> I obviously didn't properly understand the wiki page explaining it ...
> I am using cmake, but I can work with scons if the setting is easier to
>> We dropped official support for sm 1.3, partly because it is so
>> difficult to compile for. Which build system are you using? Perhaps
>> this is not properly configured.
>> > How much RAM is required to build the kernels? - that is my main
>> > Is 2.5GB limit within the OS or the build environment ?
>> > I'm using gcc 4.6.2 and CUDA 5.0
>> I don't know the number, it changes each time we add more features.
>> For sm 2.x or newer RAM usage is a lot lower than sm 1.3 so I suggest
>> just trying to find out how to configure things to avoid the issue, as
>> building sm 1.3 on 32 bit is something we don't support anymore.
> Bf-committers mailing list
> Bf-committers at blender.org
More information about the Bf-committers