Storage and Memory in Practice, Part 5: HDD Versus SSD on Cost, Power and Rebuild Time

A hard drive still stores a terabyte for less than an SSD does. IEEE’s roadmap for mass data storage says the gap in cost per GB between the two has narrowed over the past decade but is unlikely to cross, and The Memory and Storage Hierarchy places both tiers in the wider stack. Price is one of three costs a buyer filling a NASNetwork-attached storage is a storage system that gives other computers access to its files over a network. or a server pays, though, beside the power each drive draws for as long as it runs and the hours a failed drive leaves an array exposed while it rebuilds.

Part 4 read an SSD data sheet line by line. This part sets a NAS hard drive’s sheet beside two SSD sheets and works out each of those costs. Hard drive figures come from Seagate’s IronWolf Pro data sheet, and the SSD figures from Samsung’s 990 PRO data sheet and Micron’s 6600 ION launch. Rebuild times come from a test Solidigm published and from arithmetic on each drive’s rated speed, and one write rate comes from my own CrystalDiskMark runs.

What a terabyte costs

That roadmap charts the average selling price per GB of SSDs and hard drives, with data from Coughlin Associates and Objective Analysis. Read from that chart, SSDs sold for about $38 per TB in 2025 and hard drives for about $13, so a terabyte of flash cost about three times as much. Since the chart prints no values, both figures are my reading of its lines and carry a few dollars of slack either way. According to the roadmap, demand from AI data center build-outs caused a shortage of memory and storage, and NAND prices surged as a result. So the 2025 SSD figure carries that shortage in it.

What the maker took in

Seagate’s own results put a maker’s take per TB in the same range. For its fourth quarter of fiscal 2026, Seagate’s supplemental financial information reports $3,629 million of revenue against 218EB of hard drive capacity for the quarter. Divided out, that comes to about $17 for each TB. That revenue counts everything Seagate sold in the quarter, not only drives, and the quarter falls later than the chart’s 2025 figure, so the two don’t line up exactly.

What a buyer paid

Backblaze buys hard drives for its cloud storage and has published what it paid for them. Its average fell from $0.114 per GB in 2009 to $0.0144 per GB in November 2022, which is $14.40 per TB. That post ends before the shortage the roadmap describes, and it covers hard drives only.

Watts at idle and under load

Seagate rates the IronWolf Pro 32TB at 6.8W average at idle and 8.3W average while operating, and at 1.2W typical in its standby and sleep modes.

Samsung rates the 990 PRO 2TB at 55mW at idle with Autonomous Power State Transition (APST) on, and at 5.8W average while reading and 5.1W while writing, measured on an AMD Ryzen 7 5800X system. At idle the hard drive draws about 124 times as much as the SSD. Under load the gap falls to about 1.4 times, 8.3W against 5.8W.

A year at idle

Left idle for a year, the IronWolf Pro 32TB uses about 60kWh and the 990 PRO 2TB about 0.5kWh. A hard drive pool kept spinning all day pays that idle draw on every drive, while this SSD pays close to nothing between requests.

Power per terabyte in a data center

Per terabyte, the comparison turns around. Spread over 32TB, the IronWolf Pro’s draw comes to about 0.21W per TB at idle and 0.26W per TB operating, while Samsung’s 990 PRO 2TB draws about 2.9W per TB while reading and falls to 0.03W per TB only at idle.

Data center SSDs close that gap with capacity. Micron’s 6600 ION holds 245TB and draws up to 30W at maximum power, about 0.12W per TB, which is below the IronWolf Pro’s rate even at idle. Micron’s figure is a maximum and Seagate’s an average, so the comparison leans toward the hard drive. Micron says a 1EB deployment of hard drives needs 1.9 times the energy of one built on the 6600 ION, a figure from its own testing, and The Capacity Drives counts the racks each build takes.

A hard drive rebuilds no faster than it writes

A full rebuild onto one replacement drive writes every block of that drive, so it can’t finish faster than one pass at the drive’s sustained rate. Seagate rates the IronWolf Pro 32TB at up to 285MB/s, measured at the outer diameter, where the rate is highest. One full pass over 32TB at that rate takes at least 31.2 hours. At 275MB/s the 30TB model needs 30.3 hours, and at 270MB/s the 28TB model needs 28.8, so each step up in capacity adds more time than the faster rate saves.

My own bench gives the same kind of floor. In two CrystalDiskMark 8.0.4 runs at 1MiB sequential and queue depth 8, a WD Red Pro 20TB wrote at a median 271.4MB/s, so one write pass over its 20TB takes at least about 20.5 hours. Real rebuilds run longer, since the rate falls toward the inner tracks and the array keeps serving its own reads and writes. Running a NAS Part 8 works the same arithmetic for the IronWolf Pro 24TB, about 28 hours, and adds the odds of a read error along the way. Pools that rebuild only the space in use, as Synology’s Fast Repair does in Part 8, can finish sooner than a full pass, and a drive that uses shingled magnetic recording (SMR)Shingled magnetic recording overlaps each track on the one before to fit more data. Rewriting part of a band means rewriting the rest of it, which slows sustained writes. can take far longer, as Running a NAS Part 9 shows.

What measured rebuilds show

Measured rebuilds take longer than the arithmetic. The OpenZFS dRAID documentation charts observed rebuild times in a pool of 90 hard drives filled to 90% capacity, and replacing a 16TB disk with a traditional spare took about 29 hours at every group width it tested.

Busy arrays take far longer. In WorkOut, a FAST 2009 study, researchers rebuilt a disk with Linux MD’s pipeline reconstruction in an 8-disk RAID 5 set of 250GB drives, each limited to 10GB so the runs would finish. With no user I/O the rebuild took 136.4 seconds. Replaying an OLTP trace from a large financial institution stretched it to 1,121.8 seconds, about 8 times as long, and a trace from a machine running a web search engine to 9,935.6 seconds, about 73 times. Those were small, short arrays, so the factors show what user load does to a rebuild, not how long a 16TB drive would take.

An SSD rebuild is set by its software

Flash lifts the drive’s speed limit, and the RAID software takes its place. Solidigm rebuilt one 61.44TB D5-P5336 QLC drive in a RAIDRAID, a redundant array of independent disks, combines several drives into one array. Its parity levels let the array keep working when one drive fails. 5 array of nine data drives, once with Linux’s mdraid and once with Xinnor’s xiRAID Classic 4.3. With no other workload, mdraid took 53 hours 40 minutes at 322MB/s, even after Solidigm raised its rebuild speed limit from 200MB/s to 9GB/s. Xinnor’s engine took 5 hours 22 minutes at 3.18GB/s.

At mdraid’s rate the SSD rebuilt at about the speed of one hard drive, 322MB/s against the IronWolf Pro 32TB’s 285MB/s, and it took longer than the hard drive’s floor because it held more data.

Rebuilding under load

A host workload changes both results. With reads and writes running during the rebuild, xiRAID took 53 hours 53 minutes at 316MB/s. mdraid managed about 10.5MB/s, which put its rebuild at more than 67 days. Solidigm makes the drive and Xinnor makes xiRAID, so both results are vendor figures, and they still show where the time goes.

Ways to shorten a rebuild

On a traditional array every surviving drive feeds one replacement, so that one drive’s write speed sets the pace. dRAID in OpenZFS keeps spare space on every drive in the group, so a rebuild writes to all of them at once, and in the same 90-drive chart a 16TB disk rebuilt in about 7.5 hours with double parity and 8 data disks per group. Seagate’s ADAPT spreads data, parity and spare space the same way, and Seagate’s own model puts a one-drive rebuild at 55.5 hours in a RAID 6 group of 8 data and 2 parity drives against 5.3 hours across 106 drives, numbers it labels as modeled. Seagate’s model names no drive size, so its hours don’t compare with the OpenZFS chart’s.

Unraid takes a different route, which bounds the job rather than speeding it up. It stores data on individual drives instead of striping them into one volume, and no data disk can be larger than its parity disk. Unraid’s own documentation says rebuilding a disk can take several hours to more than a day for larger drives, depending on the disk’s size and system activity. Running a NAS Part 2 compares its single and dual parity.

The load can be moved aside, too. WorkOut sent user writes and popular reads to a second RAID set while the rebuild ran, which cut its OLTP-trace rebuild from 1,121.8 seconds to 203.1.

What to check before you buy

Price per TB is the one line a buyer can read straight off a shop page, and in 2025 a terabyte of flash cost about three times as much. The other two costs only show up on the data sheet and in the NAS settings.

For power, read each sheet’s idle and standby rows and count the drives that will stay spun up. The IronWolf Pro 32TB drops from 6.8W to 1.2W in standby, so whether the NAS lets its disks spin down matters as much as which disk it holds. Samsung quotes the 990 PRO’s 55mW with APST on, so check that the system keeps that state enabled.

For rebuilds, divide the drive’s capacity by its rated sustained rate to get the floor, then find out how the RAID software rebuilds. A pass over every block, a pass over used space only and an engine like the one in Solidigm’s test give very different windows on the same drives.

My take, from these numbers, is that flash earns its higher price on the data a NAS reads and writes every day, where it idles at 55mW against 6.8W a drive and can rebuild in hours with the right software. Hard drives still hold the bulk, at about a third of the price per TB, with a rebuild floor of at least 31.2 hours at 32TB that a careful pool layout has to plan around.

Current hard drive picks are in Best HDDs and SSD picks in Best SSDs, and SSDs in a NAS covers caches, tiers and all-flash pools. Next in this series is Portable SSDs and Enclosures.