diff --git a/.github/workflows/nalgebra-ci-build.yml b/.github/workflows/nalgebra-ci-build.yml
index bc2f9ca6..c00b6cbc 100644
--- a/.github/workflows/nalgebra-ci-build.yml
+++ b/.github/workflows/nalgebra-ci-build.yml
@@ -124,6 +124,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: Jimver/cuda-toolkit@v0.2.4
+        with:
+          cuda: '11.2.2'
       - name: Install nightly-2021-12-04
         uses: actions-rs/toolchain@v1
         with:
@@ -132,4 +134,6 @@ jobs:
       - uses: actions/checkout@v2
       - run: rustup target add nvptx64-nvidia-cuda
       - run: cargo build --no-default-features --features cuda
-      - run: cargo build --no-default-features --features cuda --target=nvptx64-nvidia-cuda
\ No newline at end of file
+      - run: cargo build --no-default-features --features cuda --target=nvptx64-nvidia-cuda
+        env: 
+          CUDA_ARCH: "350"
\ No newline at end of file
diff --git a/Cargo.toml b/Cargo.toml
index 15a24d41..8a3fea5c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -32,7 +32,7 @@ compare = [ "matrixcompare-core" ]
 libm    = [ "simba/libm" ]
 libm-force = [ "simba/libm_force" ]
 macros = [ "nalgebra-macros" ]
-cuda   = [ "cust", "simba/cuda" ]
+cuda   = [ "cust_core", "simba/cuda" ]
 
 # Conversion
 convert-mint = [ "mint" ]
@@ -96,9 +96,7 @@ glam017        = { package = "glam", version = "0.17", optional = true }
 glam018        = { package = "glam", version = "0.18", optional = true }
 glam019        = { package = "glam", version = "0.19", optional = true }
 glam020        = { package = "glam", version = "0.20", optional = true }
-
-[target.'cfg(not(target_os = "cuda"))'.dependencies]
-cust           = { version = "0.2", optional = true }
+cust_core      = { version = "0.1", optional = true }
 
 
 [dev-dependencies]
diff --git a/src/base/array_storage.rs b/src/base/array_storage.rs
index 6851c381..b6bd236a 100644
--- a/src/base/array_storage.rs
+++ b/src/base/array_storage.rs
@@ -27,10 +27,7 @@ use std::mem;
 /// A array-based statically sized matrix data storage.
 #[repr(transparent)]
 #[derive(Copy, Clone, PartialEq, Eq, Hash)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct ArrayStorage<T, const R: usize, const C: usize>(pub [[T; R]; C]);
 
 impl<T, const R: usize, const C: usize> ArrayStorage<T, R, C> {
diff --git a/src/base/dimension.rs b/src/base/dimension.rs
index 86006f3d..e43cb734 100644
--- a/src/base/dimension.rs
+++ b/src/base/dimension.rs
@@ -13,10 +13,7 @@ use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
 /// Dim of dynamically-sized algebraic entities.
 #[derive(Clone, Copy, Eq, PartialEq, Debug)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct Dynamic {
     value: usize,
 }
@@ -201,10 +198,7 @@ dim_ops!(
 );
 
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct Const<const R: usize>;
 
 /// Trait implemented exclusively by type-level integers.
diff --git a/src/base/matrix.rs b/src/base/matrix.rs
index bdf3a8c7..cc69c9a1 100644
--- a/src/base/matrix.rs
+++ b/src/base/matrix.rs
@@ -150,10 +150,7 @@ pub type MatrixCross<T, R1, C1, R2, C2> =
 /// some concrete types for `T` and a compatible data storage type `S`).
 #[repr(C)]
 #[derive(Clone, Copy)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct Matrix<T, R, C, S> {
     /// The data storage that contains all the matrix components. Disappointed?
     ///
diff --git a/src/base/unit.rs b/src/base/unit.rs
index 9336a5e5..bb8b56a1 100644
--- a/src/base/unit.rs
+++ b/src/base/unit.rs
@@ -21,10 +21,7 @@ use crate::{Dim, Matrix, OMatrix, RealField, Scalar, SimdComplexField, SimdRealF
 /// in their documentation, read their dedicated pages directly.
 #[repr(transparent)]
 #[derive(Clone, Hash, Copy)]
-// #[cfg_attr(
-//     all(not(target_os = "cuda"), feature = "cuda"),
-//     derive(cust::DeviceCopy)
-// )]
+// #[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct Unit<T> {
     pub(crate) value: T,
 }
@@ -102,9 +99,8 @@ mod rkyv_impl {
     }
 }
 
-#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
-unsafe impl<T: cust::memory::DeviceCopy, R, C, S> cust::memory::DeviceCopy
-    for Unit<Matrix<T, R, C, S>>
+#[cfg(feature = "cuda")]
+unsafe impl<T: cust_core::DeviceCopy, R, C, S> cust_core::DeviceCopy for Unit<Matrix<T, R, C, S>>
 where
     T: Scalar,
     R: Dim,
diff --git a/src/geometry/dual_quaternion.rs b/src/geometry/dual_quaternion.rs
index 4280668a..509b359a 100644
--- a/src/geometry/dual_quaternion.rs
+++ b/src/geometry/dual_quaternion.rs
@@ -39,10 +39,7 @@ use simba::scalar::{ClosedNeg, RealField};
 ///  See <https://github.com/dimforge/nalgebra/issues/487>
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct DualQuaternion<T> {
     /// The real component of the quaternion
     pub real: Quaternion<T>,
diff --git a/src/geometry/isometry.rs b/src/geometry/isometry.rs
index 8cdd1bfc..0179f1ff 100755
--- a/src/geometry/isometry.rs
+++ b/src/geometry/isometry.rs
@@ -50,10 +50,7 @@ use crate::geometry::{AbstractRotation, Point, Translation};
 ///
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
     feature = "serde-serialize-no-std",
diff --git a/src/geometry/orthographic.rs b/src/geometry/orthographic.rs
index 18a7852d..085ba61b 100644
--- a/src/geometry/orthographic.rs
+++ b/src/geometry/orthographic.rs
@@ -19,10 +19,7 @@ use crate::geometry::{Point3, Projective3};
 
 /// A 3D orthographic projection stored as a homogeneous 4x4 matrix.
 #[repr(C)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[derive(Copy, Clone)]
 pub struct Orthographic3<T> {
     matrix: Matrix4<T>,
diff --git a/src/geometry/perspective.rs b/src/geometry/perspective.rs
index 59b7f9f2..8ebab3e4 100644
--- a/src/geometry/perspective.rs
+++ b/src/geometry/perspective.rs
@@ -20,10 +20,7 @@ use crate::geometry::{Point3, Projective3};
 
 /// A 3D perspective projection stored as a homogeneous 4x4 matrix.
 #[repr(C)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[derive(Copy, Clone)]
 pub struct Perspective3<T> {
     matrix: Matrix4<T>,
diff --git a/src/geometry/point.rs b/src/geometry/point.rs
index b62998c3..a8d7684b 100644
--- a/src/geometry/point.rs
+++ b/src/geometry/point.rs
@@ -69,12 +69,11 @@ where
 {
 }
 
-#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
-unsafe impl<T: Scalar + cust::memory::DeviceCopy, D: DimName> cust::memory::DeviceCopy
-    for OPoint<T, D>
+#[cfg(feature = "cuda")]
+unsafe impl<T: Scalar + cust_core::DeviceCopy, D: DimName> cust_core::DeviceCopy for OPoint<T, D>
 where
     DefaultAllocator: Allocator<T, D>,
-    OVector<T, D>: cust::memory::DeviceCopy,
+    OVector<T, D>: cust_core::DeviceCopy,
 {
 }
 
diff --git a/src/geometry/quaternion.rs b/src/geometry/quaternion.rs
index 0aa7f3d3..6d29f34f 100755
--- a/src/geometry/quaternion.rs
+++ b/src/geometry/quaternion.rs
@@ -23,10 +23,7 @@ use crate::geometry::{Point3, Rotation};
 /// that may be used as a rotation.
 #[repr(C)]
 #[derive(Copy, Clone)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub struct Quaternion<T> {
     /// This quaternion as a 4D vector of coordinates in the `[ x, y, z, w ]` storage order.
     pub coords: Vector4<T>,
@@ -1045,8 +1042,8 @@ impl<T: RealField + fmt::Display> fmt::Display for Quaternion<T> {
 /// A unit quaternions. May be used to represent a rotation.
 pub type UnitQuaternion<T> = Unit<Quaternion<T>>;
 
-#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
-unsafe impl<T: cust::memory::DeviceCopy> cust::memory::DeviceCopy for UnitQuaternion<T> {}
+#[cfg(feature = "cuda")]
+unsafe impl<T: cust_core::DeviceCopy> cust_core::DeviceCopy for UnitQuaternion<T> {}
 
 impl<T: Scalar + ClosedNeg + PartialEq> PartialEq for UnitQuaternion<T> {
     #[inline]
diff --git a/src/geometry/rotation.rs b/src/geometry/rotation.rs
index 69c4a355..4dbcfb43 100755
--- a/src/geometry/rotation.rs
+++ b/src/geometry/rotation.rs
@@ -49,10 +49,7 @@ use crate::geometry::Point;
 /// * [Conversion to a matrix <span style="float:right;">`matrix`, `to_homogeneous`…</span>](#conversion-to-a-matrix)
 ///
 #[repr(C)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[derive(Copy, Clone)]
 pub struct Rotation<T, const D: usize> {
     matrix: SMatrix<T, D, D>,
diff --git a/src/geometry/scale.rs b/src/geometry/scale.rs
index 064e0075..abaeeccc 100755
--- a/src/geometry/scale.rs
+++ b/src/geometry/scale.rs
@@ -17,10 +17,7 @@ use crate::geometry::Point;
 
 /// A scale which supports non-uniform scaling.
 #[repr(C)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[derive(Copy, Clone)]
 pub struct Scale<T, const D: usize> {
     /// The scale coordinates, i.e., how much is multiplied to a point's coordinates when it is
diff --git a/src/geometry/similarity.rs b/src/geometry/similarity.rs
index 46c86f5d..9658685e 100755
--- a/src/geometry/similarity.rs
+++ b/src/geometry/similarity.rs
@@ -18,10 +18,7 @@ use crate::geometry::{AbstractRotation, Isometry, Point, Translation};
 /// A similarity, i.e., an uniform scaling, followed by a rotation, followed by a translation.
 #[repr(C)]
 #[derive(Debug, Copy, Clone)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[cfg_attr(feature = "serde-serialize-no-std", derive(Serialize, Deserialize))]
 #[cfg_attr(
     feature = "serde-serialize-no-std",
diff --git a/src/geometry/transform.rs b/src/geometry/transform.rs
index b0b5cced..2a7ca112 100755
--- a/src/geometry/transform.rs
+++ b/src/geometry/transform.rs
@@ -60,26 +60,17 @@ where
 
 /// Tag representing the most general (not necessarily inversible) `Transform` type.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub enum TGeneral {}
 
 /// Tag representing the most general inversible `Transform` type.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub enum TProjective {}
 
 /// Tag representing an affine `Transform`. Its bottom-row is equal to `(0, 0 ... 0, 1)`.
 #[derive(Debug, Copy, Clone, Hash, PartialEq, Eq)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 pub enum TAffine {}
 
 impl TCategory for TGeneral {
@@ -207,13 +198,13 @@ where
 {
 }
 
-#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
-unsafe impl<T: RealField + cust::memory::DeviceCopy, C: TCategory, const D: usize>
-    cust::memory::DeviceCopy for Transform<T, C, D>
+#[cfg(feature = "cuda")]
+unsafe impl<T: RealField + cust_core::DeviceCopy, C: TCategory, const D: usize>
+    cust_core::DeviceCopy for Transform<T, C, D>
 where
     Const<D>: DimNameAdd<U1>,
     DefaultAllocator: Allocator<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>,
-    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: cust::memory::DeviceCopy,
+    Owned<T, DimNameSum<Const<D>, U1>, DimNameSum<Const<D>, U1>>: cust_core::DeviceCopy,
 {
 }
 
diff --git a/src/geometry/translation.rs b/src/geometry/translation.rs
index b07cce20..5db46e82 100755
--- a/src/geometry/translation.rs
+++ b/src/geometry/translation.rs
@@ -17,10 +17,7 @@ use crate::geometry::Point;
 
 /// A translation.
 #[repr(C)]
-#[cfg_attr(
-    all(not(target_os = "cuda"), feature = "cuda"),
-    derive(cust::DeviceCopy)
-)]
+#[cfg_attr(feature = "cuda", derive(cust_core::DeviceCopy))]
 #[derive(Copy, Clone)]
 pub struct Translation<T, const D: usize> {
     /// The translation coordinates, i.e., how much is added to a point's coordinates when it is
diff --git a/src/geometry/unit_complex.rs b/src/geometry/unit_complex.rs
index 48405dd4..caf25493 100755
--- a/src/geometry/unit_complex.rs
+++ b/src/geometry/unit_complex.rs
@@ -31,8 +31,8 @@ use std::cmp::{Eq, PartialEq};
 /// * [Conversion to a matrix <span style="float:right;">`to_rotation_matrix`, `to_homogeneous`…</span>](#conversion-to-a-matrix)
 pub type UnitComplex<T> = Unit<Complex<T>>;
 
-#[cfg(all(not(target_os = "cuda"), feature = "cuda"))]
-unsafe impl<T: cust::memory::DeviceCopy> cust::memory::DeviceCopy for UnitComplex<T> {}
+#[cfg(feature = "cuda")]
+unsafe impl<T: cust_core::DeviceCopy> cust_core::DeviceCopy for UnitComplex<T> {}
 
 impl<T: Scalar + PartialEq> PartialEq for UnitComplex<T> {
     #[inline]