Single-core bandwidth for independent random DRAM loads

One thread performs a long stream of loads from a 16 GB array at precomputed random indices. Every load is independent of the others, each hits a different cache line, and every load misses all the way to DRAM with about 80 ns latency. The array is backed by 1 GB pages, so TLB misses are negligible, and the hardware prefetchers cannot predict the addresses. The core has a 224-entry ROB, a 72-entry load buffer, and 12 L1D line fill buffers, and the loop spends about 3 instructions per load. The DRAM channels can supply well over 50 GB/s. What bounds this single core's achieved bandwidth?

39 of 47 in Computer Architecture