CUDA: Update default architecture resolver, docs, and CI to sm_60

- PacketMath.h: Update architecture comment to reflect sm_60+ support for __ldg and native FP16.
- EigenTesting.cmake: Default EIGEN_CUDA_COMPUTE_ARCH to 60 for CUDA < 13 so the default test target exercises the newly supported floor.
- CMakeLists.txt: Update cache help text to document that empty selects sm_60 (or sm_75 on CUDA 13+).
- ci/build.linux.gitlab-ci.yml: Add 60 to EIGEN_CI_CUDA_COMPUTE_ARCH for CUDA 11.8 CI builds to provide compile-time coverage for sm_60.
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 994f2e2..d5dde5c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -715,7 +715,7 @@
   endif()
 
   set(EIGEN_CUDA_CXX_FLAGS "" CACHE STRING "Additional flags to pass to the cuda compiler.")
-  set(EIGEN_CUDA_COMPUTE_ARCH "" CACHE STRING "The CUDA compute architecture(s) to target when compiling CUDA code, as a ;-separated list of sm_XX numbers (e.g. 75;90). Empty selects sm_70, or sm_75 with CUDA 13 and later, which dropped Volta.")
+  set(EIGEN_CUDA_COMPUTE_ARCH "" CACHE STRING "The CUDA compute architecture(s) to target when compiling CUDA code, as a ;-separated list of sm_XX numbers (e.g. 75;90). Empty selects sm_60, or sm_75 with CUDA 13 and later, which dropped Pascal and Volta.")
 
   option(EIGEN_TEST_SYCL "Add Sycl support." OFF)
   if(EIGEN_TEST_SYCL)
diff --git a/Eigen/src/Core/arch/GPU/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h
index e5a6597..e65bc1f 100644
--- a/Eigen/src/Core/arch/GPU/PacketMath.h
+++ b/Eigen/src/Core/arch/GPU/PacketMath.h
@@ -19,7 +19,7 @@
 namespace internal {
 
 // Read-only data cached load (__ldg) and native FP16 arithmetic are available
-// on all supported GPU architectures (sm_70+ for CUDA, GFX906+ for HIP).
+// on all supported GPU architectures (sm_60+ for CUDA, GFX906+ for HIP).
 
 // We need to distinguish ‘clang as the CUDA compiler’ from ‘clang as the host compiler,
 // invoked by NVCC’ (e.g. on MacOS). The former needs to see both host and device implementation
diff --git a/ci/build.linux.gitlab-ci.yml b/ci/build.linux.gitlab-ci.yml
index 28c1032..953787e 100644
--- a/ci/build.linux.gitlab-ci.yml
+++ b/ci/build.linux.gitlab-ci.yml
@@ -284,10 +284,11 @@
   variables:
     # Additional flags passed to the cuda compiler.
     EIGEN_CI_CUDA_CXX_FLAGS: ""
-    # Default architecture coverage spans the GPUs used by CI: GitLab's SaaS
-    # runners are Tesla T4 (sm_75), and eigen-gpu-testing-001.gcloud is an L4
-    # (sm_89). Jobs tied to one runner pool may override this set below.
-    EIGEN_CI_CUDA_COMPUTE_ARCH: "75;89"
+    # Default architecture coverage spans the minimum supported architecture
+    # (sm_60) and the GPUs used by CI: GitLab's SaaS runners are Tesla T4
+    # (sm_75), and eigen-gpu-testing-001.gcloud is an L4 (sm_89). Jobs tied to
+    # one runner pool may override this set below.
+    EIGEN_CI_CUDA_COMPUTE_ARCH: "60;75;89"
     EIGEN_CI_BUILD_TARGET: buildtests_gpu
     EIGEN_CI_TEST_CUDA_CLANG: "off"
     EIGEN_CI_TEST_CUDA_NVC: "off"
diff --git a/cmake/EigenTesting.cmake b/cmake/EigenTesting.cmake
index ae5944b..ed47a68 100644
--- a/cmake/EigenTesting.cmake
+++ b/cmake/EigenTesting.cmake
@@ -90,14 +90,14 @@
 # ccache does not cache; expect those as misses in the statistics.
 # Resolve EIGEN_CUDA_COMPUTE_ARCH once find_package(CUDA) has set CUDA_VERSION.
 # The cache entry defaults to empty, which selects the oldest architecture the
-# toolkit still compiles for: sm_70, Eigen's documented floor (see Macros.h),
-# or sm_75 from CUDA 13 on, which dropped offline compilation for Volta.  Sets
-# the directory-scope variable the arch flags are built from; an explicit cache
-# value is used as given.
+# toolkit still compiles for: sm_60, Eigen's documented floor (see Macros.h),
+# or sm_75 from CUDA 13 on, which dropped offline compilation for Pascal and
+# Volta. Sets the directory-scope variable the arch flags are built from; an
+# explicit cache value is used as given.
 macro(ei_cuda_resolve_compute_arch)
   if("${EIGEN_CUDA_COMPUTE_ARCH}" STREQUAL "")
     if(CUDA_VERSION VERSION_LESS 13.0)
-      set(EIGEN_CUDA_COMPUTE_ARCH 70)
+      set(EIGEN_CUDA_COMPUTE_ARCH 60)
     else()
       set(EIGEN_CUDA_COMPUTE_ARCH 75)
     endif()