diff --git a/Cargo.lock b/Cargo.lock
index 5c4ca197c..8fe5771de 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1482,7 +1482,7 @@ dependencies = [
 [[package]]
 name = "rust-cuda"
 version = "0.1.0"
-source = "git+https://github.com/juntyr/rust-cuda?rev=dd9507d#dd9507d96ed34bf03a7537d62a693266ea4a8cb5"
+source = "git+https://github.com/juntyr/rust-cuda?rev=5e1534c#5e1534cf3c4bd98df88aefbfe647dcd9a519dd65"
 dependencies = [
  "const-type-layout",
  "final",
@@ -1499,7 +1499,7 @@ dependencies = [
 [[package]]
 name = "rust-cuda-derive"
 version = "0.1.0"
-source = "git+https://github.com/juntyr/rust-cuda?rev=dd9507d#dd9507d96ed34bf03a7537d62a693266ea4a8cb5"
+source = "git+https://github.com/juntyr/rust-cuda?rev=5e1534c#5e1534cf3c4bd98df88aefbfe647dcd9a519dd65"
 dependencies = [
  "proc-macro-error",
  "proc-macro2",
@@ -1510,7 +1510,7 @@ dependencies = [
 [[package]]
 name = "rust-cuda-kernel"
 version = "0.1.0"
-source = "git+https://github.com/juntyr/rust-cuda?rev=dd9507d#dd9507d96ed34bf03a7537d62a693266ea4a8cb5"
+source = "git+https://github.com/juntyr/rust-cuda?rev=5e1534c#5e1534cf3c4bd98df88aefbfe647dcd9a519dd65"
 dependencies = [
  "cargo_metadata",
  "colored",
@@ -1601,7 +1601,6 @@ dependencies = [
  "necsim-partitioning-mpi",
  "necsim-plugins-core",
  "ron",
- "rust-cuda",
  "rustcoalescence-algorithms",
  "rustcoalescence-algorithms-cuda",
  "rustcoalescence-algorithms-gillespie",
diff --git a/necsim/core/Cargo.toml b/necsim/core/Cargo.toml
index ef8b0dccc..f353fabd8 100644
--- a/necsim/core/Cargo.toml
+++ b/necsim/core/Cargo.toml
@@ -20,7 +20,7 @@ contracts = "0.6.3"
 serde = { version = "1.0", default-features = false, features = ["derive"] }
 
 [target.'cfg(target_os = "cuda")'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive"], optional = true }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive"], optional = true }
 
 [target.'cfg(not(target_os = "cuda"))'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "host"], optional = true }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "host"], optional = true }
diff --git a/necsim/core/src/landscape/extent.rs b/necsim/core/src/landscape/extent.rs
index 12d8a0219..40c1012e5 100644
--- a/necsim/core/src/landscape/extent.rs
+++ b/necsim/core/src/landscape/extent.rs
@@ -3,10 +3,12 @@ use necsim_core_bond::OffByOneU32;
 use super::Location;
 
 #[allow(clippy::module_name_repetitions, clippy::unsafe_derive_deserialize)]
-#[derive(PartialEq, Eq, Copy, Clone, Debug, serde::Deserialize, serde::Serialize, TypeLayout)]
+#[derive(PartialEq, Eq, Clone, Debug, serde::Deserialize, serde::Serialize, TypeLayout)]
+#[cfg_attr(feature = "cuda", derive(rust_cuda::lend::LendRustToCuda))]
+#[repr(C)]
+#[cfg_attr(feature = "cuda", cuda(ignore))]
 #[serde(rename = "Extent")]
 #[serde(deny_unknown_fields)]
-#[repr(C)]
 pub struct LandscapeExtent {
     x: u32,
     y: u32,
@@ -58,7 +60,7 @@ impl LandscapeExtent {
         LocationIterator {
             x: self.x,
             y: self.y,
-            extent: *self,
+            extent: self.clone(),
             first_y: true,
         }
     }
@@ -186,7 +188,7 @@ mod tests {
             LocationIterator {
                 x: 0,
                 y: 0,
-                extent,
+                extent: extent.clone(),
                 first_y: true,
             }
         );
@@ -200,7 +202,7 @@ mod tests {
             LocationIterator {
                 x: 0,
                 y: 0,
-                extent,
+                extent: extent.clone(),
                 first_y: false,
             }
         );
@@ -230,7 +232,7 @@ mod tests {
             LocationIterator {
                 x: 1386,
                 y: 6812,
-                extent,
+                extent: extent.clone(),
                 first_y: true,
             }
         );
@@ -242,7 +244,7 @@ mod tests {
             LocationIterator {
                 x: 0,
                 y: 6812,
-                extent,
+                extent: extent.clone(),
                 first_y: true,
             }
         );
@@ -255,7 +257,7 @@ mod tests {
             LocationIterator {
                 x: 1386,
                 y: 6813,
-                extent,
+                extent: extent.clone(),
                 first_y: false,
             }
         );
@@ -269,7 +271,7 @@ mod tests {
             LocationIterator {
                 x: 1386,
                 y: 0,
-                extent,
+                extent: extent.clone(),
                 first_y: false,
             }
         );
@@ -283,7 +285,7 @@ mod tests {
             LocationIterator {
                 x: 1386,
                 y: 6812,
-                extent,
+                extent: extent.clone(),
                 first_y: false,
             }
         );
diff --git a/necsim/impls/cuda/Cargo.toml b/necsim/impls/cuda/Cargo.toml
index 984ba4a50..1140bfe9f 100644
--- a/necsim/impls/cuda/Cargo.toml
+++ b/necsim/impls/cuda/Cargo.toml
@@ -15,7 +15,7 @@ contracts = "0.6.3"
 serde = { version = "1.0", default-features = false, features = ["derive"] }
 
 [target.'cfg(target_os = "cuda")'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive"] }
 
 [target.'cfg(not(target_os = "cuda"))'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "host"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "host"] }
diff --git a/necsim/impls/cuda/src/cogs/rng.rs b/necsim/impls/cuda/src/cogs/rng.rs
index 671ebb79f..0bb7feb84 100644
--- a/necsim/impls/cuda/src/cogs/rng.rs
+++ b/necsim/impls/cuda/src/cogs/rng.rs
@@ -3,7 +3,10 @@ use core::marker::PhantomData;
 use necsim_core::cogs::{MathsCore, PrimeableRng, RngCore};
 
 use const_type_layout::TypeGraphLayout;
-use rust_cuda::safety::{PortableBitSemantics, StackOnly};
+use rust_cuda::{
+    safety::{PortableBitSemantics, StackOnly},
+    utils::adapter::RustToCudaWithPortableBitCloneSemantics,
+};
 
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 
@@ -12,38 +15,45 @@ use serde::{Deserialize, Deserializer, Serialize, Serializer};
 #[cuda(free = "M", free = "R")]
 pub struct CudaRng<M: MathsCore, R>
 where
-    R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout,
+    R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout,
 {
-    inner: R,
+    #[cuda(embed)]
+    inner: RustToCudaWithPortableBitCloneSemantics<R>,
     marker: PhantomData<M>,
 }
 
-impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout> Clone
+impl<M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout + Copy> Copy
+    for CudaRng<M, R>
+{
+}
+
+#[allow(clippy::expl_impl_clone_on_copy)]
+impl<M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout> Clone
     for CudaRng<M, R>
 {
     fn clone(&self) -> Self {
         Self {
-            inner: self.inner,
+            inner: self.inner.clone(),
             marker: PhantomData::<M>,
         }
     }
 }
 
-impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout>
-    From<R> for CudaRng<M, R>
+impl<M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout> From<R>
+    for CudaRng<M, R>
 {
     #[must_use]
     #[inline]
     fn from(rng: R) -> Self {
         Self {
-            inner: rng,
+            inner: rng.into(),
             marker: PhantomData::<M>,
         }
     }
 }
 
-impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout>
-    RngCore<M> for CudaRng<M, R>
+impl<M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout> RngCore<M>
+    for CudaRng<M, R>
 {
     type Seed = <R as RngCore<M>>::Seed;
 
@@ -51,7 +61,7 @@ impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + Typ
     #[inline]
     fn from_seed(seed: Self::Seed) -> Self {
         Self {
-            inner: R::from_seed(seed),
+            inner: R::from_seed(seed).into(),
             marker: PhantomData::<M>,
         }
     }
@@ -63,10 +73,8 @@ impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + Typ
     }
 }
 
-impl<
-        M: MathsCore,
-        R: PrimeableRng<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout,
-    > PrimeableRng<M> for CudaRng<M, R>
+impl<M: MathsCore, R: PrimeableRng<M> + StackOnly + PortableBitSemantics + TypeGraphLayout>
+    PrimeableRng<M> for CudaRng<M, R>
 {
     #[inline]
     fn prime_with(&mut self, location_index: u64, time_index: u64) {
@@ -74,22 +82,19 @@ impl<
     }
 }
 
-impl<M: MathsCore, R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout>
-    Serialize for CudaRng<M, R>
+impl<M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout> Serialize
+    for CudaRng<M, R>
 {
     fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
         self.inner.serialize(serializer)
     }
 }
 
-impl<
-        'de,
-        M: MathsCore,
-        R: RngCore<M> + Copy + StackOnly + PortableBitSemantics + TypeGraphLayout,
-    > Deserialize<'de> for CudaRng<M, R>
+impl<'de, M: MathsCore, R: RngCore<M> + StackOnly + PortableBitSemantics + TypeGraphLayout>
+    Deserialize<'de> for CudaRng<M, R>
 {
     fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
-        let inner = R::deserialize(deserializer)?;
+        let inner = R::deserialize(deserializer)?.into();
 
         Ok(Self {
             inner,
diff --git a/necsim/impls/no-std/Cargo.toml b/necsim/impls/no-std/Cargo.toml
index 07f88df49..4465c45ee 100644
--- a/necsim/impls/no-std/Cargo.toml
+++ b/necsim/impls/no-std/Cargo.toml
@@ -30,7 +30,7 @@ fnv = { version = "1.0", default-features = false, features = [] }
 rand_core = "0.6"
 
 [target.'cfg(target_os = "cuda")'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "final"], optional = true }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "final"], optional = true }
 
 [target.'cfg(not(target_os = "cuda"))'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "final", "host"], optional = true }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "final", "host"], optional = true }
diff --git a/necsim/impls/no-std/src/cogs/event_sampler/independent.rs b/necsim/impls/no-std/src/cogs/event_sampler/independent.rs
index e3d51b674..17ac313d0 100644
--- a/necsim/impls/no-std/src/cogs/event_sampler/independent.rs
+++ b/necsim/impls/no-std/src/cogs/event_sampler/independent.rs
@@ -43,12 +43,7 @@ pub struct IndependentEventSampler<
     T: TurnoverRate<M, H>,
     N: SpeciationProbability<M, H>,
 > {
-    #[cfg_attr(
-        feature = "cuda",
-        cuda(
-            embed = "Option<rust_cuda::utils::adapter::RustToCudaWithPortableBitCopySemantics<SpeciationSample>>"
-        )
-    )]
+    #[cfg_attr(feature = "cuda", cuda(embed))]
     min_spec_sample: Option<SpeciationSample>,
     marker: PhantomData<(M, H, G, X, D, T, N)>,
 }
@@ -84,7 +79,7 @@ impl<
 {
     unsafe fn backup_unchecked(&self) -> Self {
         Self {
-            min_spec_sample: self.min_spec_sample,
+            min_spec_sample: self.min_spec_sample.clone(),
             marker: PhantomData::<(M, H, G, X, D, T, N)>,
         }
     }
@@ -254,7 +249,7 @@ impl<
     ) -> Option<SpeciationSample> {
         // `core::mem::replace()` would be semantically better
         //  - but `clone()` does not spill to local memory
-        let old_value = self.min_spec_sample;
+        let old_value = self.min_spec_sample.clone();
 
         self.min_spec_sample = new;
 
diff --git a/necsim/impls/no-std/src/cogs/event_sampler/tracking.rs b/necsim/impls/no-std/src/cogs/event_sampler/tracking.rs
index d541d21a4..5525256ad 100644
--- a/necsim/impls/no-std/src/cogs/event_sampler/tracking.rs
+++ b/necsim/impls/no-std/src/cogs/event_sampler/tracking.rs
@@ -26,7 +26,8 @@ pub trait MinSpeciationTrackingEventSampler<
         -> Option<SpeciationSample>;
 }
 
-#[derive(Clone, Copy, Debug, TypeLayout)]
+#[derive(Clone, Debug, TypeLayout)]
+#[cfg_attr(feature = "cuda", derive(rust_cuda::lend::LendRustToCuda))]
 #[repr(C)]
 pub struct SpeciationSample {
     speciation_sample: ClosedOpenUnitF64,
diff --git a/necsim/impls/no-std/src/cogs/habitat/in_memory.rs b/necsim/impls/no-std/src/cogs/habitat/in_memory.rs
index a1520550b..fb47d7a6d 100644
--- a/necsim/impls/no-std/src/cogs/habitat/in_memory.rs
+++ b/necsim/impls/no-std/src/cogs/habitat/in_memory.rs
@@ -21,6 +21,7 @@ pub struct InMemoryHabitat<M: MathsCore> {
     habitat: Final<Box<[u32]>>,
     #[cfg_attr(feature = "cuda", cuda(embed))]
     u64_injection: Final<Box<[u64]>>,
+    #[cfg_attr(feature = "cuda", cuda(embed))]
     extent: LandscapeExtent,
     marker: PhantomData<M>,
 }
@@ -31,7 +32,7 @@ impl<M: MathsCore> Backup for InMemoryHabitat<M> {
         Self {
             habitat: Final::new(self.habitat.clone()),
             u64_injection: Final::new(self.u64_injection.clone()),
-            extent: self.extent,
+            extent: self.extent.clone(),
             marker: PhantomData::<M>,
         }
     }
diff --git a/necsim/impls/no-std/src/cogs/habitat/non_spatial.rs b/necsim/impls/no-std/src/cogs/habitat/non_spatial.rs
index edbdf23f1..6d83fe75b 100644
--- a/necsim/impls/no-std/src/cogs/habitat/non_spatial.rs
+++ b/necsim/impls/no-std/src/cogs/habitat/non_spatial.rs
@@ -14,6 +14,7 @@ use necsim_core_bond::{OffByOneU32, OffByOneU64};
 #[cfg_attr(feature = "cuda", derive(rust_cuda::lend::LendRustToCuda))]
 #[cfg_attr(feature = "cuda", cuda(free = "M"))]
 pub struct NonSpatialHabitat<M: MathsCore> {
+    #[cfg_attr(feature = "cuda", cuda(embed))]
     extent: LandscapeExtent,
     deme: NonZeroU32,
     marker: PhantomData<M>,
@@ -58,7 +59,7 @@ impl<M: MathsCore> NonSpatialHabitat<M> {
 impl<M: MathsCore> Backup for NonSpatialHabitat<M> {
     unsafe fn backup_unchecked(&self) -> Self {
         Self {
-            extent: self.extent,
+            extent: self.extent.clone(),
             deme: self.deme,
             marker: PhantomData::<M>,
         }
diff --git a/necsim/impls/no-std/src/cogs/rng/seahash.rs b/necsim/impls/no-std/src/cogs/rng/seahash.rs
index 6487a3531..1da7d2e9a 100644
--- a/necsim/impls/no-std/src/cogs/rng/seahash.rs
+++ b/necsim/impls/no-std/src/cogs/rng/seahash.rs
@@ -17,8 +17,6 @@ pub struct SeaHash<M: MathsCore> {
     marker: PhantomData<M>,
 }
 
-impl<M: MathsCore> Copy for SeaHash<M> {}
-
 impl<M: MathsCore> RngCore<M> for SeaHash<M> {
     type Seed = [u8; 8];
 
diff --git a/necsim/impls/no-std/src/cogs/rng/wyhash.rs b/necsim/impls/no-std/src/cogs/rng/wyhash.rs
index df86272f6..eae21264c 100644
--- a/necsim/impls/no-std/src/cogs/rng/wyhash.rs
+++ b/necsim/impls/no-std/src/cogs/rng/wyhash.rs
@@ -23,8 +23,6 @@ pub struct WyHash<M: MathsCore> {
     marker: PhantomData<M>,
 }
 
-impl<M: MathsCore> Copy for WyHash<M> {}
-
 impl<M: MathsCore> RngCore<M> for WyHash<M> {
     type Seed = [u8; 8];
 
diff --git a/necsim/impls/no-std/src/decomposition/equal/area.rs b/necsim/impls/no-std/src/decomposition/equal/area.rs
index 8d61e05f2..1e418c019 100644
--- a/necsim/impls/no-std/src/decomposition/equal/area.rs
+++ b/necsim/impls/no-std/src/decomposition/equal/area.rs
@@ -12,7 +12,7 @@ impl<M: MathsCore, H: Habitat<M>> EqualDecomposition<M, H> {
     /// Returns `Ok(Self)` iff the `habitat` can be partitioned into
     ///  `subdomain.size()` by area, otherwise returns `Err(Self)`.
     pub fn area(habitat: &H, subdomain: Partition) -> Result<Self, Self> {
-        let extent = *habitat.get_extent();
+        let extent = habitat.get_extent().clone();
 
         let mut indices = Vec::with_capacity(subdomain.size().get() as usize);
 
diff --git a/necsim/impls/no-std/src/decomposition/equal/mod.rs b/necsim/impls/no-std/src/decomposition/equal/mod.rs
index d1ed79319..885c88103 100644
--- a/necsim/impls/no-std/src/decomposition/equal/mod.rs
+++ b/necsim/impls/no-std/src/decomposition/equal/mod.rs
@@ -34,7 +34,7 @@ impl<M: MathsCore, H: Habitat<M>> Backup for EqualDecomposition<M, H> {
     unsafe fn backup_unchecked(&self) -> Self {
         Self {
             subdomain: self.subdomain,
-            extent: self.extent,
+            extent: self.extent.clone(),
             morton: self.morton,
             indices: self.indices.clone(),
             _marker: PhantomData::<(M, H)>,
diff --git a/necsim/impls/no-std/src/decomposition/equal/weight.rs b/necsim/impls/no-std/src/decomposition/equal/weight.rs
index a28dbffe1..cc5ec1e86 100644
--- a/necsim/impls/no-std/src/decomposition/equal/weight.rs
+++ b/necsim/impls/no-std/src/decomposition/equal/weight.rs
@@ -12,7 +12,7 @@ impl<M: MathsCore, H: Habitat<M>> EqualDecomposition<M, H> {
     /// Returns `Ok(Self)` iff the `habitat` can be partitioned into
     ///  `subdomain.size()` by weight, otherwise returns `Err(Self)`.
     pub fn weight(habitat: &H, subdomain: Partition) -> Result<Self, Self> {
-        let extent = *habitat.get_extent();
+        let extent = habitat.get_extent().clone();
 
         let mut total_habitat = 0;
         let mut indices = Vec::with_capacity(subdomain.size().get() as usize);
diff --git a/rustcoalescence/Cargo.toml b/rustcoalescence/Cargo.toml
index 71c16c9b5..2a367eb33 100644
--- a/rustcoalescence/Cargo.toml
+++ b/rustcoalescence/Cargo.toml
@@ -14,7 +14,7 @@ necsim-partitioning-mpi = ["dep:necsim-partitioning-mpi"]
 
 rustcoalescence-algorithms-gillespie = ["dep:rustcoalescence-algorithms-gillespie"]
 rustcoalescence-algorithms-independent = ["dep:rustcoalescence-algorithms-independent"]
-rustcoalescence-algorithms-cuda = ["dep:rustcoalescence-algorithms-cuda", "dep:rust-cuda"]
+rustcoalescence-algorithms-cuda = ["dep:rustcoalescence-algorithms-cuda"]
 
 [dependencies]
 necsim-core = { path = "../necsim/core" }
@@ -34,8 +34,6 @@ rustcoalescence-algorithms-gillespie = { path = "algorithms/gillespie", optional
 rustcoalescence-algorithms-independent = { path = "algorithms/independent", optional = true }
 rustcoalescence-algorithms-cuda = { path = "algorithms/cuda", optional = true }
 
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = [], optional = true }
-
 clap = { version = "4.0", features = ["derive"] }
 anyhow = "1.0"
 serde = { version = "1.0", features = ["derive"] }
diff --git a/rustcoalescence/algorithms/cuda/Cargo.toml b/rustcoalescence/algorithms/cuda/Cargo.toml
index b51090971..e174cb6df 100644
--- a/rustcoalescence/algorithms/cuda/Cargo.toml
+++ b/rustcoalescence/algorithms/cuda/Cargo.toml
@@ -23,4 +23,4 @@ thiserror = "1.0"
 serde = { version = "1.0", features = ["derive"] }
 serde_state = "0.4"
 serde_derive_state = "0.4"
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["host"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["host"] }
diff --git a/rustcoalescence/algorithms/cuda/cpu-kernel/Cargo.toml b/rustcoalescence/algorithms/cuda/cpu-kernel/Cargo.toml
index ae5937ec8..88fe2ff9b 100644
--- a/rustcoalescence/algorithms/cuda/cpu-kernel/Cargo.toml
+++ b/rustcoalescence/algorithms/cuda/cpu-kernel/Cargo.toml
@@ -14,4 +14,4 @@ necsim-impls-no-std = { path = "../../../../necsim/impls/no-std", features = ["c
 necsim-impls-cuda = { path = "../../../../necsim/impls/cuda" }
 rustcoalescence-algorithms-cuda-gpu-kernel = { path = "../gpu-kernel" }
 
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["host"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["host"] }
diff --git a/rustcoalescence/algorithms/cuda/gpu-kernel/Cargo.toml b/rustcoalescence/algorithms/cuda/gpu-kernel/Cargo.toml
index f0a8873c0..544fe2511 100644
--- a/rustcoalescence/algorithms/cuda/gpu-kernel/Cargo.toml
+++ b/rustcoalescence/algorithms/cuda/gpu-kernel/Cargo.toml
@@ -17,7 +17,7 @@ necsim-impls-no-std = { path = "../../../../necsim/impls/no-std", features = ["c
 necsim-impls-cuda = { path = "../../../../necsim/impls/cuda" }
 
 [target.'cfg(target_os = "cuda")'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "device", "kernel"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "device", "kernel"] }
 
 [target.'cfg(not(target_os = "cuda"))'.dependencies]
-rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "dd9507d", features = ["derive", "kernel"] }
+rust-cuda = { git = "https://github.com/juntyr/rust-cuda", rev = "5e1534c", features = ["derive", "kernel"] }
diff --git a/rustcoalescence/algorithms/cuda/src/parallelisation/monolithic.rs b/rustcoalescence/algorithms/cuda/src/parallelisation/monolithic.rs
index eccfb8aba..e28cd0ef0 100644
--- a/rustcoalescence/algorithms/cuda/src/parallelisation/monolithic.rs
+++ b/rustcoalescence/algorithms/cuda/src/parallelisation/monolithic.rs
@@ -172,7 +172,6 @@ pub fn simulate<
 
     HostAndDeviceMutRef::with_new(&mut total_time_max, |total_time_max| -> Result<()> {
         HostAndDeviceMutRef::with_new(&mut total_steps_sum, |total_steps_sum| -> Result<()> {
-            // TODO: Pipeline async launches and callbacks of simulation/event analysis
             simulation.lend_to_cuda(|simulation_cuda_repr| -> Result<()> {
                 while !slow_lineages.is_empty()
                     && pause_before.map_or(true, |pause_before| level_time < pause_before)
@@ -304,6 +303,7 @@ pub fn simulate<
                         }
 
                         event_buffer = event_buffer_host_async.synchronize()?;
+                        // TODO: explore partial sorting on the GPU
                         event_buffer.report_events_unordered(&mut proxy);
 
                         proxy.local_partition().get_reporter().report_progress(
diff --git a/rustcoalescence/src/cli/simulate/dispatch/valid/algorithm_scenario.rs b/rustcoalescence/src/cli/simulate/dispatch/valid/algorithm_scenario.rs
index 6591e1d91..54715cf40 100644
--- a/rustcoalescence/src/cli/simulate/dispatch/valid/algorithm_scenario.rs
+++ b/rustcoalescence/src/cli/simulate/dispatch/valid/algorithm_scenario.rs
@@ -105,7 +105,6 @@ macro_rules! match_scenario_algorithm {
 }
 
 #[allow(clippy::too_many_arguments)]
-#[allow(clippy::too_many_lines)] // FIXME
 pub(super) fn dispatch<'p, R: Reporter, P: LocalPartition<'p, R>>(
     local_partition: P,