diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 5ad184a..a7002de 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -14,14 +14,16 @@ jobs:
         include:
           - os: ubuntu-latest
             cuda-version: "12.9"
+            cross-arch: "armv8" # test x86_64 -> arm64 cross-building, in addition to native
+            run-cross-test: false # the x86_64 runner cannot execute arm64 binaries
           - os: ubuntu-latest
             cuda-version: "13.0"
           - os: ubuntu-latest
             cuda-version: "13.2"
-          - os: ubuntu-24.04-arm
-            cuda-version: "13.2"
           - os: ubuntu-latest
             cuda-version: "13.3"
+          - os: ubuntu-24.04-arm
+            cuda-version: "13.2"
           - os: ubuntu-24.04-arm
             cuda-version: "13.3"
           # Windows:
@@ -36,12 +38,11 @@ jobs:
           - os: windows-latest
             cuda-version: "13.3"
           # 13.4.0-preview developer preview: Windows x86_64 and arm64
-          - os: windows-latest
-            cuda-version: "13.4"
-            cross-arch: "armv8" # test x86_64 -> arm64 cross-building, in addition to native
           - os: windows-11-arm
             cuda-version: "13.4"
             cross-arch: "x86_64" # test arm64 -> x86_64 cross-building, in addition to native
+            run-cross-test: true # Windows arm64 can execute x86_64 binaries
+            package-ref: "cuda-toolkit/13.4.0-preview"
 
     runs-on: ${{ matrix.os }}
 
@@ -63,13 +64,15 @@ jobs:
       - name: conan create cuda-toolkit
         run: conan create recipes/cuda-toolkit/${{ matrix.cuda-version }} && conan cache clean "*"
 
+      # test_package is skipped here: it is only built when a host compiler able to
+      # target ${{ matrix.cross-arch }} is available, in the dedicated step below
       - name: conan create cuda-toolkit (cross-building for ${{ matrix.cross-arch }})
         if: matrix.cross-arch != ''
-        run: conan create recipes/cuda-toolkit/${{ matrix.cuda-version }} -s arch=${{ matrix.cross-arch }} && conan cache clean "*"
+        run: conan create recipes/cuda-toolkit/${{ matrix.cuda-version }} -s arch=${{ matrix.cross-arch }} --test-folder="" && conan cache clean "*"
 
       - name: conan test cuda-toolkit (cross-building for ${{ matrix.cross-arch }})
-        if: matrix.cross-arch != ''
-        run: conan test recipes/cuda-toolkit/${{ matrix.cuda-version }}/test_package cuda-toolkit/13.4.0-preview -s arch=${{ matrix.cross-arch }} -c tools.cmake.cmaketoolchain:generator=Ninja
+        if: matrix.run-cross-test
+        run: conan test recipes/cuda-toolkit/${{ matrix.cuda-version }}/test_package ${{ matrix.package-ref }} -s arch=${{ matrix.cross-arch }} -c tools.cmake.cmaketoolchain:generator=Ninja
 
       - name: build cuda-samples
         if: (matrix.cuda-version == '13.2' || (matrix.cuda-version == '13.4' && startsWith(matrix.os, 'windows')))
diff --git a/recipes/cuda-toolkit/12.9/conanfile.py b/recipes/cuda-toolkit/12.9/conanfile.py
index d4feeba..0dc64e3 100644
--- a/recipes/cuda-toolkit/12.9/conanfile.py
+++ b/recipes/cuda-toolkit/12.9/conanfile.py
@@ -4,7 +4,7 @@
 
 from conan import ConanFile
 from conan.errors import ConanInvalidConfiguration
-from conan.tools.files import copy, download, get, load, rename
+from conan.tools.files import copy, download, get, load, rename, replace_in_file
 
 class CudaToolkitConan(ConanFile):
     name = "cuda-toolkit"
@@ -15,13 +15,39 @@ class CudaToolkitConan(ConanFile):
     homepage = "https://developer.nvidia.com/cuda-toolkit"
     description = "A development environment for creating high performance GPU-accelerated applications"
     settings = "os", "arch", "compiler", "build_type"
+    options = {
+        # Linux armv8 only: SBSA (Arm servers) vs Jetson/Tegra (linux-aarch64)
+        "arm_target": ["sbsa", "jetson"],
+    }
+    default_options = {
+        "arm_target": "sbsa",
+    }
+
+    def config_options(self):
+        if self.settings.os != "Linux" or self.settings.arch != "armv8":
+            del self.options.arm_target
 
     def validate(self):
         if self.settings.os not in ["Windows", "Linux"]:
             raise ConanInvalidConfiguration("Operating system not supported")
-        
+
+        # CUDA 12.9 redistributables: linux-x86_64, linux-sbsa, linux-aarch64, windows-x86_64
+        supported_archs = ["x86_64", "armv8"] if self.settings.os == "Linux" else ["x86_64"]
+        if self.settings.arch not in supported_archs:
+            raise ConanInvalidConfiguration(f"{self.settings.arch} is not supported on {self.settings.os}")
+
         # TODO: Add checks for supported compilers as per https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#id56
 
+    def _cuda_platform(self):
+        if self.settings.os == "Linux" and self.settings.arch == "armv8":
+            return "linux-aarch64" if self.options.arm_target == "jetson" else "linux-sbsa"
+        return f"{str(self.settings.os).lower()}-{self.settings.arch}"
+
+    def _targets_arch(self):
+        if self.settings.os == "Linux" and self.settings.arch == "armv8":
+            return "aarch64" if self.options.arm_target == "jetson" else "sbsa"
+        return str(self.settings.arch)
+
     def build(self):
         base_url = "https://developer.download.nvidia.com/compute/cuda/redist"
         distrib = f"redistrib_{self.version}.json"
@@ -30,8 +56,7 @@ def build(self):
         json_content = load(self, distrib)
         cuda_distrib = json.loads(json_content)
 
-        arch = "x86_64"
-        cuda_platform = f"{str(self.settings.os).lower()}-{arch}"
+        cuda_platform = self._cuda_platform()
         self.output.info(f"About to download CUDA toolkit components for {cuda_platform}")
 
         components = [
@@ -58,7 +83,8 @@ def build(self):
 
         platform_components = {
             "linux-x86_64": ["libcufile", "cuda_sandbox_dev", "cuda_opencl"],
-            "linux-aarch64": ["libcufile", "cuda_compat", "libcudla",],
+            "linux-sbsa": ["libcufile"],
+            "linux-aarch64": ["libcufile", "cuda_compat", "libcudla"],
             "windows-x86_64": ["cuda_opencl", "visual_studio_integration"],
         }
         components += platform_components.get(cuda_platform, [])
@@ -75,14 +101,20 @@ def build(self):
             get(self, url, sha256=checksum, destination="cuda", strip_root=True, keep_permissions=False)
             rename(self, f"{self.build_folder}/cuda/LICENSE", f"LICENSE_{component_name}")
 
+        if self.settings.os == "Linux":
+            # When crossbuilding, point nvcc to the target directory in the "host" package (for the target architecture),
+            # rather than have it look relative to itself (which is the current architecture of the build machine)
+            replace_in_file(self, "cuda/bin/nvcc.profile", '$(TOP)/$(_TARGET_DIR_)/', '$(CONAN_CUDA_TARGET_DIR)/')
+            replace_in_file(self, "cuda/bin/nvcc.profile", "lib$(_TARGET_SIZE_)", "lib" )
+
     def package(self):
         copy(self, "LICENSE*", src=self.build_folder, dst=os.path.join(self.package_folder, "licenses"))
 
         for folder in ["bin", "nvml", "nvvm", "compat"]:
             copy(self, "*", src=os.path.join(self.build_folder, "cuda", folder), dst=os.path.join(self.package_folder, folder))
 
-        if self.settings.os == "Linux" and not self.settings_target:
-            arch = "x86_64"
+        if self.settings.os == "Linux":
+            arch = self._targets_arch()
             targets_folder = f"targets/{arch}-linux"
             copy(self, "*", src=os.path.join(self.build_folder, "cuda", "lib"), dst=os.path.join(self.package_folder, targets_folder, "lib"))
             copy(self, "*", src=os.path.join(self.build_folder, "cuda", "include"), dst=os.path.join(self.package_folder, targets_folder, "include"))
@@ -105,7 +137,7 @@ def package(self):
                             dst=os.path.join(self.package_folder, "extras", "visual_studio_integration") )
 
     def package_id(self):
-        # only os and arch matter for this package
+        # only os, arch, and arm_target (when set) matter for this package
         self.info.settings.rm_safe("compiler")
         self.info.settings.rm_safe("build_type")
 
@@ -118,7 +150,10 @@ def package_info(self):
         self.buildenv_info.define_path("CUDACXX", os.path.join(self.package_folder, "bin", nvcc))
 
         if not self.settings_target:
+            arch = self._targets_arch() if self.settings.os == "Linux" else str(self.settings.arch)
             self.buildenv_info.define_path("CUDAToolkit_ROOT", self.package_folder)
+            if self.settings.os == "Linux":
+                self.buildenv_info.define_path("CONAN_CUDA_TARGET_DIR", os.path.join(self.package_folder, "targets", f"{arch}-linux"))
 
         if self.settings.os == "Linux":
             self.cpp_info.libdirs = ["lib64"]