Guix
November 15, 2022, 2:14pm
1
Hello nVIDIA,
Could you provide static version of the core lib cuBLAS on Windows pls? As in the case of cudart.
CUDA 11.8 comes with a huge cublasLt64_11.dll (around 530Mo!!) and cublas64_11.dll depends on it.
I am using only dgemm from cublas and I do not want to carry such a big dll with my application just for one function.
related (old) topics with no real answer from you:
Hi! Can I link statically with CUDA on Windows? Note that I’m not talking about building my static library, I mean namely static CUDA itself. Thanks!
Is this possible? I’m trying to avoid having to bundle cublas64_10.dll and cusparse64_10.dll with my application, but can’t find the equivalent .libs…
Are there any plans of releasing static versions of some of the core libs like cuBLAS on Windows?
Currently, static versions of cuBLAS are provided on Linux and OSX but not Windows. It’s been supported since CUDA 6.5 (maybe 5) but I have not seen anything at all on supporting it on Windows.
(linux flavor but about the cublasLt size)Dramatic size increase of "lightweight" cuBLASLt library from CUDA 10 to 11
Thanks,
Guillaume A.
mnicely
November 29, 2022, 12:25pm
3
@Guix at this time we have no plans to release static libraries for Windows.
amil
February 13, 2023, 10:01pm
4
I also need the static cublas library for Windows.
What is the logic that Nvidia provides the static cublas builds for Linux, but not and for Windows? Like cudart_static.lib it would simplify linkage with different CUDA toolkits.