diff --git a/cmssw-tool-conf.spec b/cmssw-tool-conf.spec
index 7e381035d16..710e64b512d 100644
--- a/cmssw-tool-conf.spec
+++ b/cmssw-tool-conf.spec
@@ -171,6 +171,10 @@ Requires: cuda-toolfile
 Requires: alpaka-toolfile
 Requires: cupla-toolfile
 
+%if "%{cmsos}" != "slc7_aarch64"
+Requires: cudnn-toolfile
+%endif
+
 %ifnarch ppc64le
 Requires: libunwind-toolfile
 Requires: igprof-toolfile
diff --git a/cudnn-toolfile.spec b/cudnn-toolfile.spec
new file mode 100644
index 00000000000..2aed88d3eb4
--- /dev/null
+++ b/cudnn-toolfile.spec
@@ -0,0 +1,22 @@
+### RPM external cudnn-toolfile 1.0
+Requires: cudnn
+%prep
+
+%build
+
+%install
+mkdir -p %i/etc/scram.d
+cat << \EOF_TOOLFILE >%i/etc/scram.d/cudnn.xml
+<tool name="cudnn" version="@TOOL_VERSION@">
+  <info url="https://docs.nvidia.com/deeplearning/cudnn/index.html"/>
+  <lib name="cudnn"/>
+  <client>
+    <environment name="CUDNN_BASE" default="@TOOL_ROOT@"/>
+    <environment name="INCLUDE" default="$CUDNN_BASE/include"/>
+    <environment name="LIBDIR" default="$CUDNN_BASE/lib64"/>
+  </client>
+  <use name="cuda"/>
+</tool>
+EOF_TOOLFILE
+
+## IMPORT scram-tools-post
diff --git a/cudnn.spec b/cudnn.spec
new file mode 100644
index 00000000000..778b2f5f22e
--- /dev/null
+++ b/cudnn.spec
@@ -0,0 +1,37 @@
+### RPM external cudnn 8.1.1.33
+## INITENV +PATH LD_LIBRARY_PATH %i/lib64
+
+%define cudaver 11.2
+%define cudnnver_maj %(echo %{realversion} | cut -f1,2,3 -d.)
+
+%ifarch x86_64
+Source: https://developer.download.nvidia.com/compute/redist/cudnn/v%{cudnnver_maj}/cudnn-%{cudaver}-linux-x64-v%{realversion}.tgz
+%endif
+%ifarch ppc64le
+Source: https://developer.download.nvidia.com/compute/redist/cudnn/v%{cudnnver_maj}/cudnn-%{cudaver}-linux-ppc64le-v%{realversion}.tgz
+%endif
+%ifarch aarch64
+Source: https://developer.download.nvidia.com/compute/redist/cudnn/v%{cudnnver_maj}/cudnn-%{cudaver}-linux-aarch64sbsa-v%{realversion}.tgz
+%endif
+Requires: cuda
+
+%prep
+%setup -n cuda
+
+if [ "${CUDA_VERSION%.*}" != %{cudaver} ]; then
+    echo 'Incompatible CUDA version in cudnn.spec!'
+    exit 1
+fi
+
+%build
+
+%install
+%ifarch ppc64le
+rm -f %_builddir/cuda/targets/ppc64le-linux/lib/*.a
+mv %_builddir/cuda/targets/ppc64le-linux/lib %{i}/lib64
+mv %_builddir/cuda/targets/ppc64le-linux/* %{i}/
+%else
+rm -f %_builddir/cuda/lib64/*.a
+mv %_builddir/cuda/* %{i}/
+%endif
+
diff --git a/onnxruntime-toolfile.spec b/onnxruntime-toolfile.spec
index cac1d127593..116f28ab527 100644
--- a/onnxruntime-toolfile.spec
+++ b/onnxruntime-toolfile.spec
@@ -16,6 +16,10 @@ cat << \EOF_TOOLFILE >%i/etc/scram.d/onnxruntime.xml
     <environment name="LIBDIR" default="$ONNXRUNTIME_BASE/lib"/>
   </client>
   <use name="protobuf"/>
+%if "%{cmsos}" != "slc7_aarch64"
+  <use name="cuda"/>
+  <use name="cudnn"/>
+%endif
   <runtime name="MLAS_DYNAMIC_CPU_ARCH" value="0"/>
 </tool>
 EOF_TOOLFILE
diff --git a/onnxruntime.spec b/onnxruntime.spec
index d13b02ebc27..89de3830034 100644
--- a/onnxruntime.spec
+++ b/onnxruntime.spec
@@ -7,6 +7,9 @@ Source: git+https://github.com/%{github_user}/%{n}.git?obj=%{branch}/%{tag}&expo
 
 BuildRequires: cmake ninja
 Requires: protobuf py3-numpy py2-wheel py2-onnx zlib libpng py2-pybind11
+%if "%{cmsos}" != "slc7_aarch64"
+Requires: cuda cudnn
+%endif
 
 %prep
 %setup -q -n %{n}-%{realversion}
@@ -14,6 +17,12 @@ Requires: protobuf py3-numpy py2-wheel py2-onnx zlib libpng py2-pybind11
 %build
 rm -rf ../build; mkdir ../build; cd ../build
 
+%if "%{cmsos}" != "slc7_aarch64"
+USE_CUDA=ON
+%else
+USE_CUDA=OFF
+%endif
+
 cmake ../%{n}-%{realversion}/cmake -GNinja \
    -DPYTHON_EXECUTABLE=${PYTHON3_ROOT}/bin/python3 \
    -DCMAKE_BUILD_TYPE=Release \
@@ -21,7 +30,10 @@ cmake ../%{n}-%{realversion}/cmake -GNinja \
    -DCMAKE_INSTALL_LIBDIR=lib \
    -Donnxruntime_ENABLE_PYTHON=ON \
    -Donnxruntime_BUILD_SHARED_LIB=ON \
-   -Donnxruntime_USE_CUDA=OFF \
+   -Donnxruntime_USE_CUDA=${USE_CUDA} \
+   -Donnxruntime_CUDA_VERSION="${CUDA_VERSION}" \
+   -Donnxruntime_CUDA_HOME="${CUDA_ROOT}" \
+   -Donnxruntime_CUDNN_HOME="${CUDNN_ROOT}" \
    -Donnxruntime_BUILD_CSHARP=OFF \
    -Donnxruntime_USE_EIGEN_FOR_BLAS=ON \
    -Donnxruntime_USE_OPENBLAS=OFF \
@@ -39,6 +51,9 @@ cmake ../%{n}-%{realversion}/cmake -GNinja \
    -Donnxruntime_DISABLE_CONTRIB_OPS=OFF \
    -Donnxruntime_USE_PREINSTALLED_PROTOBUF=ON \
    -Donnxruntime_PREFER_SYSTEM_LIB=ON \
+   -DCMAKE_CUDA_FLAGS="-cudart shared" \
+   -DCMAKE_CUDA_RUNTIME_LIBRARY=Shared \
+   -DCMAKE_TRY_COMPILE_PLATFORM_VARIABLES="CMAKE_CUDA_RUNTIME_LIBRARY" \
    -DCMAKE_PREFIX_PATH="${ZLIB_ROOT};${LIBPNG_ROOT};${PROTOBUF_ROOT};${PY2_PYBIND11_ROOT}"
 
 ninja -v %{makeprocesses}