diff --git a/libcudacxx/include/cuda/__device/arch_id.h b/libcudacxx/include/cuda/__device/arch_id.h
index 65f4d6f455c..1e37c7fd65e 100644
--- a/libcudacxx/include/cuda/__device/arch_id.h
+++ b/libcudacxx/include/cuda/__device/arch_id.h
@@ -23,6 +23,7 @@
 
 #include <cuda/__device/compute_capability.h>
 #include <cuda/__fwd/devices.h>
+#include <cuda/std/__fwd/format.h>
 #include <cuda/std/__type_traits/always_false.h>
 #include <cuda/std/__utility/to_underlying.h>
 #include <cuda/std/array>
@@ -142,6 +143,40 @@ enum class arch_id : int
 
 _CCCL_END_NAMESPACE_CUDA
 
+#if __cpp_lib_format >= 201907L
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _CharT>
+struct formatter<::cuda::arch_id, _CharT> : private formatter<::cuda::compute_capability, _CharT>
+{
+  template <class _ParseCtx>
+  _CCCL_HOST_API constexpr auto parse(_ParseCtx& __ctx)
+  {
+    return __ctx.begin();
+  }
+
+  template <class _FmtCtx>
+  _CCCL_HOST_API auto format(const ::cuda::arch_id& __arch, _FmtCtx& __ctx) const
+  {
+    auto __it = __ctx.out();
+    *__it++   = _CharT{'s'};
+    *__it++   = _CharT{'m'};
+    *__it++   = _CharT{'_'};
+    __ctx.advance_to(__it);
+    __it = formatter<::cuda::compute_capability, _CharT>::format(::cuda::compute_capability{__arch}, __ctx);
+    if (::cuda::__is_specific_arch(__arch))
+    {
+      *__it++ = _CharT{'a'};
+    }
+    return __it;
+  }
+};
+
+_CCCL_END_NAMESPACE_STD
+#endif // __cpp_lib_format >= 201907L
+
+// todo: specialize cuda::std::formatter for cuda::arch_id
+
 #if _CCCL_CUDA_COMPILATION()
 
 _CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
diff --git a/libcudacxx/include/cuda/__device/compute_capability.h b/libcudacxx/include/cuda/__device/compute_capability.h
index 7e73e986034..6ef414b5976 100644
--- a/libcudacxx/include/cuda/__device/compute_capability.h
+++ b/libcudacxx/include/cuda/__device/compute_capability.h
@@ -22,6 +22,7 @@
 #endif // no system header
 
 #include <cuda/__fwd/devices.h>
+#include <cuda/std/__fwd/format.h>
 #include <cuda/std/__utility/to_underlying.h>
 
 #include <cuda/std/__cccl/prologue.h>
@@ -172,6 +173,30 @@ class compute_capability
 
 _CCCL_END_NAMESPACE_CUDA
 
+#if __cpp_lib_format >= 201907L
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _CharT>
+struct formatter<::cuda::compute_capability, _CharT> : private formatter<int, _CharT>
+{
+  template <class _ParseCtx>
+  _CCCL_HOST_API constexpr auto parse(_ParseCtx& __ctx)
+  {
+    return __ctx.begin();
+  }
+
+  template <class _FmtCtx>
+  _CCCL_HOST_API auto format(const ::cuda::compute_capability& __cc, _FmtCtx& __ctx) const
+  {
+    return formatter<int, _CharT>::format(__cc.get(), __ctx);
+  }
+};
+
+_CCCL_END_NAMESPACE_STD
+#endif // __cpp_lib_format >= 201907L
+
+// todo: specialize cuda::std::formatter for cuda::compute_capability
+
 #if _CCCL_CUDA_COMPILATION()
 
 _CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
diff --git a/libcudacxx/include/cuda/std/__fwd/format.h b/libcudacxx/include/cuda/std/__fwd/format.h
index 98e9253c2aa..b9371dc5a26 100644
--- a/libcudacxx/include/cuda/std/__fwd/format.h
+++ b/libcudacxx/include/cuda/std/__fwd/format.h
@@ -24,6 +24,17 @@
 
 #include <cuda/std/__cccl/prologue.h>
 
+#if __cpp_lib_format >= 201907L
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class, class>
+struct formatter;
+
+_CCCL_END_NAMESPACE_STD
+
+#endif // __cpp_lib_format >= 201907L
+
 _CCCL_BEGIN_NAMESPACE_CUDA_STD
 
 template <class _CharT>
diff --git a/libcudacxx/test/libcudacxx/cuda/ccclrt/device/arch_id_fmt.pass.cpp b/libcudacxx/test/libcudacxx/cuda/ccclrt/device/arch_id_fmt.pass.cpp
new file mode 100644
index 00000000000..b1b86b7cf89
--- /dev/null
+++ b/libcudacxx/test/libcudacxx/cuda/ccclrt/device/arch_id_fmt.pass.cpp
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/devices>
+
+#if __cpp_lib_format >= 201907L
+#  include <format>
+#endif // __cpp_lib_format >= 201907L
+
+#include "literal.h"
+
+#if __cpp_lib_format >= 201907L
+template <class C>
+void test()
+{
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_60) == TEST_STRLIT(C, "sm_60"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_61) == TEST_STRLIT(C, "sm_61"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_62) == TEST_STRLIT(C, "sm_62"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_70) == TEST_STRLIT(C, "sm_70"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_75) == TEST_STRLIT(C, "sm_75"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_80) == TEST_STRLIT(C, "sm_80"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_86) == TEST_STRLIT(C, "sm_86"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_87) == TEST_STRLIT(C, "sm_87"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_88) == TEST_STRLIT(C, "sm_88"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_89) == TEST_STRLIT(C, "sm_89"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_90) == TEST_STRLIT(C, "sm_90"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_100) == TEST_STRLIT(C, "sm_100"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_103) == TEST_STRLIT(C, "sm_103"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_110) == TEST_STRLIT(C, "sm_110"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_120) == TEST_STRLIT(C, "sm_120"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_121) == TEST_STRLIT(C, "sm_121"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_90a) == TEST_STRLIT(C, "sm_90a"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_100a) == TEST_STRLIT(C, "sm_100a"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_103a) == TEST_STRLIT(C, "sm_103a"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_110a) == TEST_STRLIT(C, "sm_110a"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_120a) == TEST_STRLIT(C, "sm_120a"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::arch_id::sm_121a) == TEST_STRLIT(C, "sm_121a"));
+}
+
+void test()
+{
+  test<char>();
+  test<wchar_t>();
+}
+#endif // __cpp_lib_format >= 201907L
+
+int main(int, char**)
+{
+#if __cpp_lib_format >= 201907L
+  NV_IF_TARGET(NV_IS_HOST, (test();))
+#endif // __cpp_lib_format >= 201907L
+  return 0;
+}
diff --git a/libcudacxx/test/libcudacxx/cuda/ccclrt/device/compute_capability_fmt.pass.cpp b/libcudacxx/test/libcudacxx/cuda/ccclrt/device/compute_capability_fmt.pass.cpp
new file mode 100644
index 00000000000..8039c4c20e2
--- /dev/null
+++ b/libcudacxx/test/libcudacxx/cuda/ccclrt/device/compute_capability_fmt.pass.cpp
@@ -0,0 +1,55 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/devices>
+
+#if __cpp_lib_format >= 201907L
+#  include <format>
+#endif // __cpp_lib_format >= 201907L
+
+#include "literal.h"
+
+#if __cpp_lib_format >= 201907L
+template <class C>
+void test()
+{
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{0}) == TEST_STRLIT(C, "0"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{60}) == TEST_STRLIT(C, "60"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{61}) == TEST_STRLIT(C, "61"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{62}) == TEST_STRLIT(C, "62"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{70}) == TEST_STRLIT(C, "70"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{75}) == TEST_STRLIT(C, "75"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{80}) == TEST_STRLIT(C, "80"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{86}) == TEST_STRLIT(C, "86"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{87}) == TEST_STRLIT(C, "87"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{88}) == TEST_STRLIT(C, "88"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{89}) == TEST_STRLIT(C, "89"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{90}) == TEST_STRLIT(C, "90"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{100}) == TEST_STRLIT(C, "100"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{103}) == TEST_STRLIT(C, "103"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{110}) == TEST_STRLIT(C, "110"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{120}) == TEST_STRLIT(C, "120"));
+  assert(std::format(TEST_STRLIT(C, "{}"), cuda::compute_capability{121}) == TEST_STRLIT(C, "121"));
+}
+
+void test()
+{
+  test<char>();
+  test<wchar_t>();
+}
+#endif // __cpp_lib_format >= 201907L
+
+int main(int, char**)
+{
+#if __cpp_lib_format >= 201907L
+  NV_IF_TARGET(NV_IS_HOST, (test();))
+#endif // __cpp_lib_format >= 201907L
+  return 0;
+}