diff --git a/testing/cuda_timer.cu b/testing/cuda_timer.cu
index 751bcd8..0c85a3b 100644
--- a/testing/cuda_timer.cu
+++ b/testing/cuda_timer.cu
@@ -3,28 +3,11 @@
 #include <nvbench/cuda_stream.cuh>
 #include <nvbench/types.cuh>
 
+#include "sleep_kernel.cuh"
 #include "test_asserts.cuh"
 
 #include <fmt/format.h>
 
-#include <cuda/std/chrono>
-
-#include <cuda_runtime.h>
-
-__global__ void sleep_kernel(double seconds)
-{
-  const auto start = cuda::std::chrono::high_resolution_clock::now();
-  const auto ns    = cuda::std::chrono::nanoseconds(
-    static_cast<nvbench::int64_t>(seconds * 1000 * 1000 * 1000));
-  const auto finish = start + ns;
-
-  auto now = cuda::std::chrono::high_resolution_clock::now();
-  while (now < finish)
-  {
-    now = cuda::std::chrono::high_resolution_clock::now();
-  }
-}
-
 void test_basic(cudaStream_t time_stream,
                 cudaStream_t exec_stream,
                 bool expected)
diff --git a/testing/sleep_kernel.cuh b/testing/sleep_kernel.cuh
new file mode 100644
index 0000000..2d18f87
--- /dev/null
+++ b/testing/sleep_kernel.cuh
@@ -0,0 +1,19 @@
+#pragma once
+
+#include <cuda/std/chrono>
+
+#include <cuda_runtime.h>
+
+__global__ void sleep_kernel(double seconds)
+{
+  const auto start = cuda::std::chrono::high_resolution_clock::now();
+  const auto ns    = cuda::std::chrono::nanoseconds(
+    static_cast<nvbench::int64_t>(seconds * 1000 * 1000 * 1000));
+  const auto finish = start + ns;
+
+  auto now = cuda::std::chrono::high_resolution_clock::now();
+  while (now < finish)
+  {
+    now = cuda::std::chrono::high_resolution_clock::now();
+  }
+}