diff --git a/dev-cpp/eigen/eigen-3.4.9999.ebuild b/dev-cpp/eigen/eigen-3.4.9999.ebuild
index 7aeca7b126824..21ac6c0a075a1 100644
--- a/dev-cpp/eigen/eigen-3.4.9999.ebuild
+++ b/dev-cpp/eigen/eigen-3.4.9999.ebuild
@@ -114,6 +114,10 @@ REQUIRED_USE="
 	)
 	test? (
 		|| ( ${IUSE_TEST_BACKENDS[*]} )
+		cuda-clang? (
+			!openmp
+			${LLVM_REQUIRED_USE}
+		)
 	)
 "
 
@@ -193,7 +197,8 @@ DEPEND="
 			)
 			cuda-clang? (
 				$(llvm_gen_dep '
-					llvm-core/clang:${LLVM_SLOT}[llvm_targets_NVPTX]
+					llvm-core/clang:${LLVM_SLOT}
+					llvm-runtimes/clang-runtime:${LLVM_SLOT}[llvm_targets_NVPTX,offload,openmp]
 				')
 			)
 		)
@@ -239,16 +244,13 @@ cuda_set_CUDAHOSTCXX() {
 			break
 		fi
 		CUDAHOSTCXX_test="$(
-			dirname "$(
-				realpath "$(
-					which "${compiler}-$(echo "${version}" | grep -oP "(?<=${package}-)[0-9]*")"
-				)"
-			)"
+			which "${compiler}-$(echo "${version}" | grep -oP "(?<=${package}-)[0-9]*")"
 		)"
 		version="<${version}"
 	do ! echo "int main(){}" | nvcc "-ccbin ${CUDAHOSTCXX_test}" - -x cu &>/dev/null; done
 
 	export CUDAHOSTCXX
+	echo "${CUDAHOSTCXX}"
 }
 
 pkg_pretend() {
@@ -368,6 +370,14 @@ src_configure() {
 			-DEIGEN_DEBUG_ASSERTS="$(usex debug)" # Enable advanced debugging of assertions
 			-DEIGEN_SPLIT_LARGE_TESTS="yes" # Split large tests into smaller executables
 
+			-DEIGEN_TEST_CUDA="$(usex cuda)" # Enable CUDA support in unit tests
+			-DEIGEN_TEST_CUDA_CLANG="$(usex cuda "$(usex cuda-clang)")" # Use clang instead of nvcc to compile the CUDA tests
+
+			-DEIGEN_TEST_HIP="$(usex hip)" # Add HIP support.
+
+			# -DEIGEN_TEST_SYCL="$(usex sycl)" # Add Sycl support.
+			# -DEIGEN_SYCL_TRISYCL="no" # Use the triSYCL Sycl implementation (ComputeCPP by default).
+
 			$(cmake_use_find_package adolc Adolc)
 			$(cmake_use_find_package boost Boost)
 			$(cmake_use_find_package cholmod CHOLMOD)
@@ -437,48 +447,59 @@ src_configure() {
 			)
 		fi
 
-		mycmakeargs+=(
-			-DEIGEN_TEST_CUDA="$(usex cuda)" # Enable CUDA support in unit tests
-			-DEIGEN_TEST_CUDA_CLANG="$(usex cuda "$(usex cuda-clang)")" # Use clang instead of nvcc to compile the CUDA tests
-
-			-DEIGEN_TEST_HIP="$(usex hip)" # Add HIP support.
-
-			# -DEIGEN_TEST_SYCL=no # Add Sycl support.
-			# -DEIGEN_SYCL_TRISYCL=no # Use the triSYCL Sycl implementation (ComputeCPP by default).
-		)
-
 		if use cuda; then
 			cuda_add_sandbox -w
+
 			if use cuda-clang; then
 				local llvm_prefix
 				llvm_prefix="$(get_llvm_prefix -b)"
-				export CC="${llvm_prefix}/bin/clang"
-				export CXX="${llvm_prefix}/bin/clang++"
-				export LIBRARY_PATH="${ESYSROOT}/usr/$(get_libdir)"
+
+				# NVCCFLAGS gets injected into CMAKE_CXX_FLAGS, which means we need to use clang as gcc will fail with
+				# error: unrecognized command-line option
+				if ! tc-is-clang; then
+					export CC="${CHOST}-clang-${LLVM_SLOT}"
+					export CXX="${CHOST}-clang++-${LLVM_SLOT}"
+				fi
+
+				NVCCFLAGS="${NVCCFLAGS:+${NVCCFLAGS} }--libomptarget-nvptx-bc-path=${llvm_prefix}/$(get_libdir)/nvptx64-nvidia-cuda/libomptarget-nvptx.bc"
+
 				mycmakeargs+=(
-					-DCUDA_HOST_COMPILER="${llvm_prefix}/bin/clang++"
+					-DCUDA_HOST_COMPILER="${CHOST}-clang++-${LLVM_SLOT}"
 				)
 			else
 				cuda_set_CUDAHOSTCXX
+
 				mycmakeargs+=(
 					-DCUDA_HOST_COMPILER="${CUDAHOSTCXX}"
 				)
+				if [[ -v CUDACXX ]]; then
+					mycmakeargs+=(
+						-DCUDA_NVCC_EXECUTABLE="${CUDACXX}"
+						-DCMAKE_POLICY_DEFAULT_CMP0146="OLD" # FindCUDA
+					)
+				fi
 			fi
+
 			if [[ "${CUDA_VERBOSE}" == true ]]; then
 				mycmakeargs+=(
-					-DCUDA_VERBOSE_BUILD=yes
+					-DCUDA_VERBOSE_BUILD="yes"
 				)
 				NVCCFLAGS+=" -v"
 			fi
 
-			export CUDAFLAGS="${NVCCFLAGS}"
-
-			[[ -z "${CUDAARCHS}" ]] && einfo "trying to determine host CUDAARCHS"
-			: "${CUDAARCHS:=$(__nvcc_device_query)}"
-			export CUDAARCHS
+			if [[ -v CUDAARCHS ]]; then
+				mycmakeargs+=(
+					# TODO this needs to be lowest first for eigen
+					-DEIGEN_CUDA_COMPUTE_ARCH="${CUDAARCHS}"
+				)
+			fi
 
+			# CUDAFLAGS is used by cmake
+			# NVCCFLAGS is used by cuda.eclass
 			mycmakeargs+=(
-				-DEIGEN_CUDA_COMPUTE_ARCH="${CUDAARCHS}"
+				-DEIGEN_CUDA_CXX_FLAGS="${NVCCFLAGS}"
+
+				-DCUDA_USE_STATIC_CUDA_RUNTIME="no"
 			)
 		fi
 
diff --git a/dev-cpp/eigen/eigen-9999.ebuild b/dev-cpp/eigen/eigen-9999.ebuild
index a7da9aab56530..5eccde3fc263e 100644
--- a/dev-cpp/eigen/eigen-9999.ebuild
+++ b/dev-cpp/eigen/eigen-9999.ebuild
@@ -114,6 +114,10 @@ REQUIRED_USE="
 	)
 	test? (
 		|| ( ${IUSE_TEST_BACKENDS[*]} )
+		cuda-clang? (
+			!openmp
+			${LLVM_REQUIRED_USE}
+		)
 	)
 "
 
@@ -193,7 +197,8 @@ DEPEND="
 			)
 			cuda-clang? (
 				$(llvm_gen_dep '
-					llvm-core/clang:${LLVM_SLOT}[llvm_targets_NVPTX]
+					llvm-core/clang:${LLVM_SLOT}
+					llvm-runtimes/clang-runtime:${LLVM_SLOT}[llvm_targets_NVPTX,offload,openmp]
 				')
 			)
 		)
@@ -236,16 +241,13 @@ cuda_set_CUDAHOSTCXX() {
 			break
 		fi
 		CUDAHOSTCXX_test="$(
-			dirname "$(
-				realpath "$(
-					which "${compiler}-$(echo "${version}" | grep -oP "(?<=${package}-)[0-9]*")"
-				)"
-			)"
+			which "${compiler}-$(echo "${version}" | grep -oP "(?<=${package}-)[0-9]*")"
 		)"
 		version="<${version}"
 	do ! echo "int main(){}" | nvcc "-ccbin ${CUDAHOSTCXX_test}" - -x cu &>/dev/null; done
 
 	export CUDAHOSTCXX
+	echo "${CUDAHOSTCXX}"
 }
 
 pkg_pretend() {
@@ -367,6 +369,14 @@ src_configure() {
 			-DEIGEN_DEBUG_ASSERTS="$(usex debug)" # Enable advanced debugging of assertions
 			-DEIGEN_SPLIT_LARGE_TESTS="yes" # Split large tests into smaller executables
 
+			-DEIGEN_TEST_CUDA="$(usex cuda)" # Enable CUDA support in unit tests
+			-DEIGEN_TEST_CUDA_CLANG="$(usex cuda "$(usex cuda-clang)")" # Use clang instead of nvcc to compile the CUDA tests
+
+			-DEIGEN_TEST_HIP="$(usex hip)" # Add HIP support.
+
+			# -DEIGEN_TEST_SYCL="$(usex sycl)" # Add Sycl support.
+			# -DEIGEN_SYCL_TRISYCL="no" # Use the triSYCL Sycl implementation (ComputeCPP by default).
+
 			$(cmake_use_find_package adolc Adolc)
 			$(cmake_use_find_package boost Boost)
 			$(cmake_use_find_package cholmod CHOLMOD)
@@ -439,48 +449,59 @@ src_configure() {
 			)
 		fi
 
-		mycmakeargs+=(
-			-DEIGEN_TEST_CUDA="$(usex cuda)" # Enable CUDA support in unit tests
-			-DEIGEN_TEST_CUDA_CLANG="$(usex cuda "$(usex cuda-clang)")" # Use clang instead of nvcc to compile the CUDA tests
-
-			-DEIGEN_TEST_HIP="$(usex hip)" # Add HIP support.
-
-			# -DEIGEN_TEST_SYCL=no # Add Sycl support.
-			# -DEIGEN_SYCL_TRISYCL=no # Use the triSYCL Sycl implementation (ComputeCPP by default).
-		)
-
 		if use cuda; then
 			cuda_add_sandbox -w
+
 			if use cuda-clang; then
 				local llvm_prefix
 				llvm_prefix="$(get_llvm_prefix -b)"
-				export CC="${llvm_prefix}/bin/clang"
-				export CXX="${llvm_prefix}/bin/clang++"
-				export LIBRARY_PATH="${ESYSROOT}/usr/$(get_libdir)"
+
+				# NVCCFLAGS gets injected into CMAKE_CXX_FLAGS, which means we need to use clang as gcc will fail with
+				# error: unrecognized command-line option
+				if ! tc-is-clang; then
+					export CC="${CHOST}-clang-${LLVM_SLOT}"
+					export CXX="${CHOST}-clang++-${LLVM_SLOT}"
+				fi
+
+				NVCCFLAGS="${NVCCFLAGS:+${NVCCFLAGS} }--libomptarget-nvptx-bc-path=${llvm_prefix}/$(get_libdir)/nvptx64-nvidia-cuda/libomptarget-nvptx.bc"
+
 				mycmakeargs+=(
-					-DCUDA_HOST_COMPILER="${llvm_prefix}/bin/clang++"
+					-DCUDA_HOST_COMPILER="${CHOST}-clang++-${LLVM_SLOT}"
 				)
 			else
 				cuda_set_CUDAHOSTCXX
+
 				mycmakeargs+=(
 					-DCUDA_HOST_COMPILER="${CUDAHOSTCXX}"
 				)
+				if [[ -v CUDACXX ]]; then
+					mycmakeargs+=(
+						-DCUDA_NVCC_EXECUTABLE="${CUDACXX}"
+						-DCMAKE_POLICY_DEFAULT_CMP0146="OLD" # FindCUDA
+					)
+				fi
 			fi
+
 			if [[ "${CUDA_VERBOSE}" == true ]]; then
 				mycmakeargs+=(
-					-DCUDA_VERBOSE_BUILD=yes
+					-DCUDA_VERBOSE_BUILD="yes"
 				)
 				NVCCFLAGS+=" -v"
 			fi
 
-			export CUDAFLAGS="${NVCCFLAGS}"
-
-			[[ -z "${CUDAARCHS}" ]] && einfo "trying to determine host CUDAARCHS"
-			: "${CUDAARCHS:=$(__nvcc_device_query)}"
-			export CUDAARCHS
+			if [[ -v CUDAARCHS ]]; then
+				mycmakeargs+=(
+					# TODO this needs to be lowest first for eigen
+					-DEIGEN_CUDA_COMPUTE_ARCH="${CUDAARCHS}"
+				)
+			fi
 
+			# CUDAFLAGS is used by cmake
+			# NVCCFLAGS is used by cuda.eclass
 			mycmakeargs+=(
-				-DEIGEN_CUDA_COMPUTE_ARCH="${CUDAARCHS}"
+				-DEIGEN_CUDA_CXX_FLAGS="${NVCCFLAGS}"
+
+				-DCUDA_USE_STATIC_CUDA_RUNTIME="no"
 			)
 		fi
 
diff --git a/dev-cpp/eigen/metadata.xml b/dev-cpp/eigen/metadata.xml
index 0bc2f84f498ce..17014deef36d9 100644
--- a/dev-cpp/eigen/metadata.xml
+++ b/dev-cpp/eigen/metadata.xml
@@ -17,14 +17,14 @@
 		providing optimal speed with GCC.
 	</longdescription>
 	<use>
-		<flag name="clang">Use clang for cuda units compilation</flag>
-		<flag name="clang-cuda">Use clang for cuda units compilation</flag>
-		<flag name="hip">Enable HIP GPU computing support</flag>
 		<flag name="adolc">Add test support for Adolc (<pkg>sci-libs/adolc</pkg>)</flag>
 		<flag name="blas">Build the Eigen Blas library</flag>
 		<flag name="boost">Add test support for Boost.Multiprecision (<pkg>dev-libs/boost</pkg>)</flag>
 		<flag name="cholmod">Add test support for CHOLMOD (<pkg>sci-libs/cholmod</pkg>)</flag>
+		<flag name="clang">Use clang for cuda units compilation</flag>
+		<flag name="cuda-clang">Use clang for cuda units compilation</flag>
 		<flag name="fftw">Add test support for fftw (<pkg>sci-libs/fftw</pkg>)</flag>
+		<flag name="hip">Enable HIP GPU computing support</flag>
 		<flag name="klu">Add test support for KLU (<pkg>sci-libs/klu</pkg>)</flag>
 		<flag name="lapack">Build the included Eigen LAPACK library</flag>
 		<flag name="mathjax">Use MathJax for rendering math in HTML docs (dev-libs/mathjax)</flag>