Fresh-out-of-box Data Explains Nothing
A device straight from the factory or freshly purged is in the FOB state: every physical block is empty, writes land directly in free blocks, and garbage collection has almost nothing to do. Random write measured there approaches the theoretical ceiling of media and channels and can exceed the long-run operating level several times over. The number is not wrong in itself; it simply describes a transient that does not last. Selecting on FOB data amounts to planning capacity and performance on the assumption that the drive is never filled.
Purge and Preconditioning Are Two Different Things
A disciplined flow begins with purge, using format or sanitize to return the device to a known starting point independent of prior workload so that runs remain comparable. Preconditioning follows in two stages. The workload-independent stage writes sequentially over twice the user capacity to guarantee every physical block has been written. The workload-dependent stage then runs the target workload continuously until the device settles into the block distribution and reclamation rhythm that workload produces. Neither stage substitutes for the other — the block state left by sequential writing is nothing like the one left by random writing.
Steady State Is Not a Flat-looking Curve
Steady state needs a decidable criterion. The usual practice takes a run of consecutive measurement rounds as an observation window, checks that every data point in the window falls within ten percent of the window average, and additionally fits a straight line through the window data, requiring that the fitted line's values at the two ends differ by no more than ten percent of the average. Only when both conditions hold is steady state declared, and the number of rounds required to get there must be recorded, since that figure itself reflects how quickly the reclamation mechanism converges.
Lock the Reporting Conditions
A comparable report has to state its conditions in full: preconditioning method and duration, the read/write mix and block size matrix, queue depth and thread count, ambient temperature and cooling, and over-provisioning setting. The most common failure is mix switching. Changing the read/write ratio changes reclamation pressure, so steady state under the previous mix no longer holds under the new one; the device must be preconditioned again or the matrix traversed continuously in the prescribed order, otherwise numbers in different cells cannot be compared. Latency data likewise needs its collection conditions stated — latency at QD1 and latency at full queue describe entirely different questions and cannot be quoted interchangeably.
Which Basis the Internal Targets Use
For the CORE 7 series, the planned 7,400 MB/s sequential read, 5,500 MB/s sequential write and 1,500 KIOPS 4KB random read are performance targets; the 300 KIOPS and 600 KIOPS random write targets for the J750 and J770 are defined on a steady-state basis; and the 65µs and 10µs read and write latency targets correspond to QD1 conditions. All of these are design targets today, validation will work through them item by item using the flow above, and the test records will retain the full preconditioning history and the steady-state determination behind each figure.