Copy-and-kernel pairs issued on one stream versus two

Assume pinned host memory and a GPU with separate copy and compute engines. Compare the two launch sequences below. Which operations can actually execute concurrently?

kernel.cu
// Sequence A: everything on the default stream
cudaMemcpyAsync(d_a, h_a, bytes, cudaMemcpyHostToDevice);
kernelA<<<blocks, threads>>>(d_a);
cudaMemcpyAsync(d_b, h_b, bytes, cudaMemcpyHostToDevice);
kernelB<<<blocks, threads>>>(d_b);

// Sequence B: two independent streams
cudaMemcpyAsync(d_a, h_a, bytes, cudaMemcpyHostToDevice, stream1);
kernelA<<<blocks, threads, 0, stream1>>>(d_a);
cudaMemcpyAsync(d_b, h_b, bytes, cudaMemcpyHostToDevice, stream2);
kernelB<<<blocks, threads, 0, stream2>>>(d_b);

24 of 46 in GPU / CUDA