Timing a kernel with std::chrono around the launch statement

d_x holds 2^28 floats (1 GiB) in device memory. The kernel reads and rewrites every element once, so it moves about 2 GiB of DRAM traffic per run on a GPU with roughly 2 TB/s of memory bandwidth. The program has already done a warm-up launch. The printed interval is consistently about 7 microseconds. What explains the number?

bench.cu
__global__ void scale(float* x, size_t n, float s) {
    size_t i = blockIdx.x * (size_t)blockDim.x + threadIdx.x;
    if (i < n) x[i] *= s;
}

const size_t n = size_t(1) << 28;
auto t0 = std::chrono::steady_clock::now();
scale<<<(n + 255) / 256, 256>>>(d_x, n, 2.0f);
auto t1 = std::chrono::steady_clock::now();
printf("%.1f us\n",
       std::chrono::duration<double, std::micro>(t1 - t0).count());

31 of 46 in GPU / CUDA