In case this helps someone else with an RTX 3060 who wants to use CUDA 13.0, here's my modified PKGBUILD:
@@ -14,7 +14,7 @@
makedepends=(
'cmake'
'cuda'
-# 'cudnn'
+ 'cudnn'
'gcc14'
'git'
'intel-oneapi-mkl'
@@ -96,8 +96,9 @@
-DOPENBLAS_INCLUDE_DIR='/usr/include/openblas' \
-DWITH_RUY='ON' \
-DWITH_CUDA='ON' \
+ -DWITH_CUDNN='ON' \
-DCUDA_DYNAMIC_LOADING='ON' \
- -DCUDA_ARCH_LIST='Common' \
+ -DCUDA_NVCC_FLAGS="-gencode arch=compute_86,code=sm_86" \
-DCMAKE_POLICY_VERSION_MINIMUM='3.5' \
-DENABLE_CPU_DISPATCH='OFF' \
-Wno-dev
The arch=compute_86
refers to the compute capability of the 3060, which is 8.6.
Pinned Comments
yochananmarqos commented on 2024-10-24 19:03 (UTC)
PSA: I just enabled OPENBLAS, but I had to disable both MKL and DNNL to do it. Please let me know how that goes.