Zero-filling a 1 GB buffer with regular vs. streaming stores

On a typical x86-64 core, a routine fills a 1 GB buffer, far larger than the last-level cache, with zeros using 32-byte AVX stores. Version A uses ordinary stores (_mm256_store_si256) to normal write-back memory. Version B uses non-temporal streaming stores (_mm256_stream_si256) followed by one _mm_sfence(). Both write every byte of every cache line, in order. Measured at the memory controller, Version A moves close to twice as many bytes over the DRAM bus as Version B. Why?

main.cpp
// Version A
for (size_t i = 0; i < n; i += 32)
    _mm256_store_si256((__m256i*)(buf + i), zero);

// Version B
for (size_t i = 0; i < n; i += 32)
    _mm256_stream_si256((__m256i*)(buf + i), zero);
_mm_sfence();

35 of 47 in Computer Architecture