The problem

How much does request order matter when accesses target different rows in the same bank?

What current research shows

DRAM uses banks and row buffers. Research on parallelism-aware batch scheduling shows that requests can create bank and row-buffer conflicts and can destroy another thread's bank-level parallelism, reducing both fairness and throughput.

Where the evidence stops

The mapping from addresses to channels, ranks, banks, and rows is hardware-dependent and often requires measurement or documentation.

What Valen Systems is testing

Measure row-buffer hit rate, bank conflicts, queue depth, stall time, fairness, and work per joule under controlled access patterns.

Sources

Mutlu and Moscibroda: parallelism-aware DRAM schedulingStaged memory scheduling for heterogeneous systems