Fifty kernel passes over a 1 GB cudaHostAllocMapped buffer

On an A100 attached by PCIe 4.0 x16 (about 25 GB/s in practice; HBM is about 1.5 TB/s), an engineer avoids an explicit copy by allocating the input as mapped pinned memory and passing the device view straight to the kernel. Each launch reads the whole 1 GB input once, in a fully coalesced pattern. The engineer expected the loop to take tens of milliseconds; it takes about 2 seconds. What explains it, and what is the fix?

relax.cu
float* h_in;
cudaHostAlloc(&h_in, bytes, cudaHostAllocMapped);   // 1 GB, filled on the host
float* d_view;
cudaHostGetDevicePointer(&d_view, h_in, 0);

for (int iter = 0; iter < 50; ++iter)
    relax<<<grid, block>>>(d_view, d_out, n);       // reads all of d_view
cudaDeviceSynchronize();

44 of 46 in GPU / CUDA