Storage and Memory in Practice, Part 9: Used Enterprise SSDs for a Home Lab

A used enterprise SSD usually arrives with most of its endurance left, because its first owner rarely wrote what the rating allowed. ServeTheHome recorded the power-on hours and data written for 1,347 used data center SSDs it bought over about eight years. The majority had been written at under 1 drive writes per day (DWPD)Drive writes per day is the number of times the full capacity can be written each day for the warranty period, which equals TBW divided by capacity and days.. Most of the drives carried ratings of 1 to 3 DWPD. Wear isn’t usually the problem. The problems are the connector, the firmware, the data the last owner left, and the airflow the drive was built for.

SSDs for Builders is Part 3 of this series and covered new drives in M.2 slots. This part covers the drives data centers retire. It explains what a server SSD adds, how to read what the first owner left in its SMART log, how a U.2 drive reaches a desktop, and what to check before the first write. Prices and picks stay off this page because a used drive’s price moves with each listing. Best SSDs carries the new-drive alternatives. Numbers here come from ServeTheHome’s two studies, maker pages, the NVMe specification, and my own logs and bench runs.

What a server SSD adds

Server SSDs are built for a bay that never idles and a power supply that can fail. For a home lab the biggest difference is power-loss protection (PLPPower loss protection uses stored energy to let a drive finish writing buffered data and metadata when power fails, so acknowledged writes survive.). Samsung’s white paper says its enterprise SSDs get full PLP from tantalum capacitors, and that client drives normally don’t have it because the capacitors cost too much. Solidigm’s D7-P5520 brief adds a firmware check, built into the drive, that confirms data was saved accurately after power returns.

In Figure 9.1, the charge gauge falls as the cache empties into NAND. Flash Foundations Part 6 explains how the capacitors let the controller finish its pending writes after a power cut. Endurance is the second difference. ServeTheHome calls the D7-P5520 the 1 DWPD mixed-use level of drives, and it puts today’s drives at over 10PB of write endurance even at 1 DWPD, against the much smaller drives in its 2016 used set.

One of my bench runs shows a steady rate. In Iometer 1.1.0, after a six-hour random write precondition, my Intel DC P4510 2TB wrote 4KiB random data at a queue depth (QD) of 256 at 88,440 IOPS over the last hour and 88,758 IOPS at the QD256 sweep step that followed. That flat line is what steady stateSteady state is the condition where performance has stopped changing because the drive is full of data and garbage collection runs continuously. looks like, and Benchmarking 201, Part 3 has the full runs.

Reading what the first owner left

Ask the seller for the drive’s log before you buy, or read it the day the drive arrives. On an NVMe drive the SMART health log holds the fields that matter, and smartctl -a /dev/nvme0 prints them without writing to the drive. Six fields in the NVMe log definition are worth reading.

• Percentage Used is the vendor’s estimate of life spent. A value of 100 means the estimated endurance is consumed, and the value can go higher.
• Data Units Written counts in thousands of 512-byte units, so each reported unit is 512,000 bytes, and the controller converts to 512-byte units even on a drive formatted with larger blocks.
• Power On Hours may leave out time the controller spent in a non-operational power state.
• Unsafe Shutdowns counts power losses that arrived with no shutdown notice.
• Media and Data Integrity Errors counts errors the controller couldn’t recover, and Critical Warning sets bit 0 when the spare blocks fall below their threshold.

Dividing data written by power-on days and by capacity gives what the first owner really used, and ServeTheHome gets its own figures the same way. It is the rating’s formula turned around, DWPD = TBW / 365 / warranty / capacity, from the SNIA endurance white paper, and How to Read an SSD Spec Sheet covers the rating itself. On the wear-out myth page my 4TB system drive shows 42.9TB written over 10,726 power-on hours, which is 447 days. That is about 96GB a day, or 0.024 DWPD, and the drive reports Percentage UsedPercentage Used is a vendor estimate of drive life consumed, based on writes and wear. It can pass 100%, and the drive may keep working. of 1%.

Figure 9.2 ends on 0.024 DWPD for that drive, with the log’s Percentage Used of 1% as the cross-check.

ServeTheHome’s 2016 study of second-hand drives reached 412 of them and could read data written on 234. Of the second-hand drives it bought, 88% had under 150TBW, and most of the drives with readable counters sat below 0.1 DWPD. In the 2024 set of 1,347 drives most workloads ran under 1 DWPD. Only the Intel 750 ran near its rating, and it is a consumer SSD in data center form factors. The largest group of drives held 3.84TB, and its highest write rate was under 1 DWPD.

Three cautions come with those studies. ServeTheHome discloses that Solidigm sponsored its 2024 study and says it collected the data well before Solidigm existed. Hours also pile up with no writes on a drive that sat powered on before it went into production, which lowers its average. And some drives report no history at all, because the Dell OEM drives ServeTheHome bought report neither data written nor power-on hours. You can’t judge a drive with a blank counter, so pass on it or ask for another way to see its history.

Getting a U.2 drive into a desktop

The U.2 form keeps the 2.5-inch shape and an SFF-8639 connector, and it comes in 7mm and 15mm heights. U.3 keeps the same connector but lets one bay take NVMe, SAS, or SATA drives, and Flash Foundations Part 1 lists the other server form factors. Desktop boards have no such bay, so the drive needs an adapter.

StarTech’s PEX4SFF8639 is a PCIe 4.0 x4 card that holds a U.2 or U.3 drive up to 15mm tall in an x4, x8, or x16 slot. It takes NVMe drives only, and SATA and SAS aren’t supported. Its vented bracket and board cutout are there, by StarTech’s account, to avoid thermal throttling. Its M2E4SFF8643 card goes the other way and moves a U.2 drive onto an M.2 x4 host interface through an SFF-8643 cable connector.

Figure 9.3 shows the two routes and the one that fails. Each U.2 drive takes four lanes, so the shared-lane and chipset-slot rules in Part 3 apply to it as they do to an M.2 drive. Server drives also expect server airflow. Read the drive’s temperature in the SMART log after a long write run, and add a fan before you trust a 15mm drive in a tower that has none.

Before the first write

ServeTheHome’s advice from 2016 still holds: test every used drive before it goes into production, because a used drive can arrive dead or with vendor-locked firmware. Two SanDisk Lightning Ascend Gen II 800GB drives it bought failed burn-in, probably from firmware problems.

Firmware is the second check. Cisco’s field notice FN74253 covers a firmware defect in certain NVMe and SATA SSDs in Cisco UCS servers that can fail a drive through improper voltage handling. It offers no workaround after a failure and names a firmware upgrade as the fix. Its list includes Solidigm D7-P5520 2.5-inch U.2 15mm parts sold under Cisco part numbers, and the D7-P5520 is the drive ServeTheHome reviewed. Some used drives may still carry old firmware, so read the maker’s or the OEM’s release notes for the exact model before you put data on it.

Third comes the data. In 2016, ServeTheHome found that 121 of its second-hand drives, over 29%, still had accessible data on them, and by 2024 the number had gone down significantly. Treat every drive as unwiped and erase it yourself. nvme sanitize sends the NVMe Sanitize command, and nvme format with a secure erase setting erases all user data wherever it sits, cache and deallocated blocks included. Send back any drive that refuses either command.

In Figure 9.4 a failed check takes a drive out of the chain, and only a drive that clears all three gets a write.

Last, fill the drive once and read it back, using the jobs in Benchmarking 201, Part 1, and save the SMART log from the day it arrived. Read the log again a month into your own workload to see what you add, and keep the first owner’s counters as the baseline.

Which drive to look for

This guide reads ServeTheHome’s numbers as an argument for capacity over endurance. In its 1,347-drive set, the smaller drives tended to see much higher drive writes per day than the larger ones. The highest use among the 400GB drives came just short of their rated endurance. An 800GB drive rated 10+ DWPD had been written at under 3 DWPD, and the 3.84TB group peaked under 1 DWPD. At 3.84TB and 1 DWPD, a used drive has better odds of arriving with most of its life left than a small high-endurance one, and it holds more.

Figure 9.5 sets the highest write rates ServeTheHome saw against the rated levels.

Job type is the exception. ServeTheHome found that some applications needed higher-endurance drives, and a ZFS write log or a write cache is that kind of job, so check the log’s counters on any drive meant for it. SSDs in a NAS covers which role a cache, a log, and a pool each ask of a drive. For everything else, read the log twice, once at arrival and once a month in, and the difference tells you the one number no listing can: how fast your own workload spends what the first owner left. Part 8, Data Center Form Factors, covers the rest of the server form factors once it is live.