HDD workload ratings: count reads, check the clock

16 October 2026 / Source review and synthetic arithmetic

A drive's workload rating is not a countdown to failure. Before comparing a measured number with a specification, identify the traffic, the clock and the device. HDD annualized reads plus writes and SSD write endurance are not interchangeable measures.

The denominator is powered-on time

Seagate defines annualized workload from lifetime host reads plus writes, scaled to 8,760 power-on hours. Western Digital also defines workload as user data transferred to or from the HDD and uses recorded power-on hours. Neither formula is simply bytes transferred since January 1. Use the limit and qualifications for the actual model, not an old product-class table. [1] [3]

Annualized TB/year = (read TB + written TB) × 8,760 / power-on hours

This is a normalization convention, not a forecast that the drive will run for another year. One TB here is 10¹² bytes. Never put a calendar-time denominator under a power-on-time label.

The same bytes can produce a different rate

Hesela arithmetic examples, not measured drives. Reads and writes cover the same interval as the power-on hours. Values are decimal TB.

Host transfer volume and annualized rate under declared power-on hours
Illustration	Read TB	Written TB	Power-on h	Annualized TB/year
Same volume, continuous power	100	20	8,760	120
Same volume, half the power-on hours	100	20	4,380	240
52 complete 20 TB host reads	1,040	0	8,760	1,040

The first two rows transfer identical volumes. Halving powered-on time doubles the normalized rate. Adding idle powered-on time lowers this arithmetic average without undoing prior activity. That is a denominator effect, not evidence that deliberately idling a drive restores reliability.

The third row assumes exactly 52 complete, host-issued reads of a 20 TB device during 8,760 powered-on hours. It produces 1,040 TB/year before other traffic. This is not a claim that every filesystem scrub reads full device capacity: allocation, redundancy, implementation and caches change device traffic. Device-internal self-tests are not automatically host transfers. Keep the verification scope and accounting boundary separate.

Do not disable verification based on this illustration. It exposes a planning question: can the hardware support both useful work and required integrity checks? The table does not calculate a safe scan frequency or the consequences of missing corruption.

A lifetime average can hide a recent change

Consider invented counters at two observations of the same drive. At the first, reads are 900 TB, writes 100 TB and powered-on time 43,800 hours. At the second, they are 980 TB, 120 TB and 44,520 hours. The lifetime average rises from 200 to 216.44 TB/year, while the recent 720-hour interval annualizes to 1,216.67 TB/year.

The recent value uses differences for all three counters. Dividing a month's bytes by lifetime hours mixes windows and understates activity. A short-window rate is an operational diagnostic, not a replacement for a manufacturer's stated lifetime calculation or evidence of a warranty breach. Save both snapshots and label the window.

Reject zero elapsed power-on time, counter decreases, device replacement or unknown units. Reset and rollover need explicit treatment. A missing read counter is unknown, not zero. Application bytes cannot substitute for per-drive transfers without accounting for the intervening storage stack.

Why HDD workload is not SSD TBW

Western Digital's January 2023 white paper distinguishes workload-conditioned HDD reliability from NAND write endurance. It describes thermal fly-height control as changing head-to-media spacing during active transfers. This is the manufacturer's mechanism explanation, not a published per-drive failure equation. It also separates maximum workload from the conditions underlying a quoted MTBF.[2, pp. 6-8]

SSD endurance ratings describe host writes under a specified workload; write amplification and retention requirements matter. TBW does not count reads as extra host writes or count raw NAND program/erase cycles. A rating is not an exact failure instant. [2, pp. 4-5]

Kingston's KC600 footnote 4 ties TBW to JESD219A, while footnote 3 describes a warranty condition using time or its SSD Life Remaining indicator. This rejects a universal definition equating TBW with warranty coverage; actual terms remain product-specific. [5]

Multiplying an HDD annual rating by five years creates an arithmetic quantity, not a transferable lifetime endurance allowance. It cannot justify unlimited bursts, a remaining-life percentage or an equal-footing comparison with SSD TBW.

What the field paper establishes

Pinheiro, Weber and Barroso's FAST 2007 study collected detailed observations from December 2005 through August 2006 on more than 100,000 consumer-grade drives in Google's datacenters. Its failure event was a repair replacement, excluding upgrades; the timestamp was replacement time. The utilization analysis used weekly read/write bandwidth and within-model categories, not modern vendor workload-limit exceedance.[4, sections 2.2, 2.3 and 3.3]

The relationship between activity and replacement differed across age groups and was weaker than a simple monotonic rule would suggest. The study is observational, with historical hardware and undisclosed model-level breakdowns. It does not calibrate today's TB/year ratings, prove activity harmless, or supply a multiplier for predicting failure after exceeding a rating. No raw-data reanalysis or replication was performed here. [4]

Hesela's inference is narrower: vendor operating conditions, observed activity and failure outcomes need separate fields. Translating one into another requires matching populations, event definitions and a validated model. See MTTF versus field replacement rates.

Reproduce the arithmetic

The table and counter example use the same dependency-free JavaScript module; the JSON result records inputs, units and assumptions. Tests cover normal values, counter resets, zero time, invalid numbers and overflow. No confidence intervals are attached to invented inputs.

For an actual assessment, retain model and firmware, manual revision, counter source and units, observation times, power-on hours, read/write totals, temperature conditions and maintenance policy. These calculations check accounting consistency; they do not replace the reliability specification.

Corpus: workload rate limit, TBW and DWPD.

Primary sources

Reviewed 16 October 2026. Manufacturer claims and historical observations retain their original scope. Original figures and source PDFs are not republished.