A vectorized scale loop whose buffers differ by 4 KB plus 64 bytes

This loop is compiled with AVX2 (8 floats, 32 bytes, per vector iteration) and runs on an Intel Skylake core. src and dst are separate 8 KB buffers, and the function is called millions of times so both stay resident in L1D. In configuration A, (dst - src) % 4096 == 64 bytes. In configuration B, the same allocations are shifted so that (dst - src) % 4096 == 2048. Configuration A is consistently and substantially slower, with no extra cache misses. What is the cause?

main.cpp
void scale(float* __restrict dst, const float* __restrict src,
           size_t n, float k) {
    for (size_t i = 0; i < n; ++i)
        dst[i] = src[i] * k;   // vector load src, multiply, vector store dst
}

40 of 47 in Computer Architecture