Under a Microsecond per Command
NVMe allows up to 65,535 SQ/CQ pairs, but how many a drive can actually saturate depends on how many independent command pipelines the controller has. The CORE 7 J750 and J770 run PCIe 4.0 x4 with a planned 4KB random read target of 1,500 KIOPS, leaving under 0.7 µs of end-to-end budget per command on average. At that scale any cross-core shared resource — a lock on the FTL mapping table, a completion queue region polled by several cores — becomes a throughput ceiling rather than something to tune later.
Three Real Cost Centers
First, doorbells: every batch the host submits costs a doorbell register write, and at high IOPS those MMIO writes consume upstream transactions, so firmware must fetch SQ entries in batches instead of reacting one at a time. Second, mapping-table locality: the LBA distribution of a random workload sets the mapping-page hit rate, and an affinity scheme that scatters adjacent LBAs across cores destroys that locality. Third, the completion path: interrupt coalescing cuts host CPU interrupt counts sharply, but each coalescing window adds latency.
QD1 Latency Versus Deep-Queue Throughput
J750/J770 target 65 µs read and 10 µs write average latency at QD1, the write figure resting on a write-back cache backed by power loss protection. Those numbers only hold with the coalescing window nearly closed. Rather than seeking one setting for everything, latency-sensitive metadata and journal traffic gets a very short window, while bandwidth workloads at 7,400 MB/s sequential read and 5,500 MB/s sequential write tolerate more aggressive coalescing. Firmware exposes both as runtime-switchable QoS profiles, not compile-time constants.
Where Validation Stands
Scheduling work runs on the internal JH-E4U-C engineering platform, both products remain at the product-definition/EVT stage, and every figure above is a planning target rather than a measured commitment. The current priority is making queue count, affinity policy and coalescing window orthogonal, configurable items so measurements from different server platforms can feed straight back into the default parameter table without a firmware rebuild each time.