How much power does an enterprise U.2 NVMe SSD draw, and can I make it idle lower?

1. 1.Measure at the wall. In the NVMe structures we read, no field reports a drive's live draw: each power state carries a maximum and, optionally, a typical figure, so the threads fall back on a meter. That absence is our reading of the documents opened, not a vendor statement.

   Measure at the wall (plug-in meter or metered PDU/UPS) with the system idle, then power down, remove only the drive (or add a second identical drive), and measure again

   ReadingThe difference is the drive's idle draw plus anything it changes elsewhere, such as the CPU package's idle states or an adapter. Pull other cards one at a time to separate them: in one ServeTheHome thread the owner of a P4510 that seemed to idle at about 20 W removed an Intel X550-T2 NIC, recovered about 10 W, and now suspects the drive uses about 5 W; the thread holds no measurement with the drive removed. Repeat under a sustained write load for the active figure, and compare both with the maker's typical figures rather than the ps 0 mp ceiling.
2. 2.Check ASPM on the drive's link and on every other card. A later reply in the same ServeTheHome thread pointed at cards without ASPM keeping the CPU out of its idle states, which a drive gets blamed for.

   sudo lspci -vv -s <bus:dev.fn> | grep -E 'LnkCap|LnkCtl'

   ReadingLnkCap says whether a link can use ASPM and LnkCtl whether it does now; the drop-out page linked below gives the full readings. For the power question, a NIC or HBA whose link never uses ASPM is the next card to pull for the wall test.

1. 1.Print the power-state table the drive reports about itself. It is the drive's own statement, so read it before any forum figure.

   sudo nvme id-ctrl -H /dev/nvme0 | grep -E 'npss|apsta|Autonomous Power|wctemp|cctemp|^ps|rwt:|active_power_workload'

   Readingnpss counts from zero: 0 means the drive has one power state, PS0. apsta : 0x1 means APST is supported, 0 not supported; the decoded line under it reads 'Autonomous Power State Transitions Supported' or 'Not Supported'. Each 'ps N : mp:X.XXW' line is one state, marked operational or non-operational: mp is that state's maximum, a ceiling rather than a measurement, and enlat and exlat are microseconds. Where the drive reports them, idle_power and active_power are its own typical figures, and active_power_workload names the workload behind the active one; '-' means not reported. wctemp and cctemp are in Kelvin; subtract 273 for °C. The labels are nvme-cli 2.8's and can differ in other versions. Threads show enterprise drives sitting in PS0 or without APST (a P4510, Kingston U.2 drives, Samsung 983 DCT drives); we found no maker document listing which models expose non-operational states, so this output is the only place to check yours.
2. 2.Ask the drive whether the kernel programmed idle transitions into it. Autonomous Power State Transition is feature 0x0c.

   sudo nvme get-feature /dev/nvme0 -f 0x0c -H

   ReadingWith -H, nvme-cli prints 'Autonomous Power State Transition Enable (APSTE): Enabled' or 'Disabled', then every entry of the transition table, each with an ITPT (idle time before the transition, in milliseconds) and an ITPS (the state it moves to). Enabled with at least one entry whose ITPT is above 0 ms means the drive drops into that idle state on its own. Disabled means APST was turned off, for example with default_ps_max_latency_us set to 0. Enabled with every entry at ITPT 0 ms and ITPS 0 means the kernel found no non-operational state within its latency limit and programmed an empty table, so the drive gets no autonomous idle saving. If id-ctrl showed apsta 0, the kernel does not program APST at all. The drop-out page linked below decodes the rest of this output.
3. 3.Check the kernel's latency limit, which decides which non-operational states APST may use.

   cat /sys/module/nvme_core/parameters/default_ps_max_latency_us

   Reading100000 microseconds is the default and 0 means APST is off on this boot. A non-operational state slower than the limit is never used, so a drive whose only non-operational states are slower idles as if it had none. When no state qualifies, because the drive lists no non-operational state or none within the latency limit, the kernel logs 'APST enabled but no non-operational states are available' at debug level only, so a default dmesg does not show it. If the value is 0 because APST was turned off to stop a drive dropping out, read the drop-out page linked below before changing it.

1. 1.Read the power state the drive is in now. Power Management is feature 2.

   sudo nvme get-feature /dev/nvme0 -f 2 -H

   Reading'Power State (PS): 0' means the drive is in its highest-power state, where threads report a P4510 and Kingston U.2 drives sitting at idle. 'Workload Hint (WH)' prints alongside it. Add -s 3 to see whether this drive can save and change the feature.
2. 2.If id-ctrl listed more than one operational state, ask for a higher-numbered one. The command writes a drive setting, not data: nothing on the drive is erased. It caps power and, going by the descriptor's own relative throughput and latency ranks, performance with it. Without -s the setting lasts only until the next reset; with -s it is saved across resets on drives that support saving, and a maker's management tool may override it either way. Set it back to 0 to undo.

   sudo nvme set-feature /dev/nvme0 -f 2 -V <N>

   ReadingThe drive accepts only a state it lists; any other value fails with Invalid Field in Command. Read the result back with the get-feature command above, then repeat the wall measurement and your own workload to see what the cap cost. Threads also mention the Intel and Solidigm tool's power-governor setting as a route to the same end; we did not read its documentation, so this page gives no modes or figures for it.

1. 1.Add up the drives from published figures, one line per model, with the source beside each number. Kioxia lists the CD8-R and CD8-V (2.5-inch, PCIe 4.0) at 11-20 W typical active and 5 W typical Ready, and the CD8P-R (2.5-inch, PCIe 5.0) at 18-24 W active and 5-5.5 W Ready; the larger active figures sit under the larger capacities in Kioxia's table columns, but the text does not state that pairing, so it is our reading. Dell's R760 table for rear NVMe drives (2 or 4 rear 2.5-inch drives with a rear fan in 24 x 2.5-inch SAS or 12 x 3.5-inch chassis) lists, among them: Samsung PM9A3 7.68TB 14 W, Hynix PE8010 (7.68, 3.84 and 1.92TB) 17 W, Kioxia CD7 and CD8 15.36TB 19 W, Samsung PM1733a 15.36TB 19.7 W and PM1735a 12.8TB 19.8 W, Intel P5520 15.36TB 20 W, PE8110 (7.68, 3.84 and 1.92TB), PS1010 15.36TB and PS1030 12.8TB at 20 W, Samsung PM1733 15.36TB 22 W, a 'Redtail' 7.68TB 24.5 W and Kioxia CM7 30.72TB 25 W. Those are Dell's ratings for its own qualified parts in that table; we have not read a Samsung, Micron, Solidigm or WD data sheet for this page, so no maker figure for those drives appears here. A smaller form factor is not a lower idle figure by itself: Kioxia lists the E1.S XD7P and the E3.S CD8P-R at the same 5 W typical Ready as the 2.5-inch CD8-R, with 16-20 W and 18-23 W active.
2. 2.Treat the total as a sum of sheet figures and nothing more, then add the rest of the box: the NIC, the HBA and the CPUs. We have not read an SFF-8639, SFF-TA-1001 or EDSFF document stating a connector or per-slot power limit, so this page gives no such figure, and we have not read a source for converting watts into heat units, so it gives no heat figure either.
3. 3.Check the server's own thermal table, because the drives you fit change the room temperature the server is rated for. Dell's R760 table for rear NVMe drives allows each listed drive a maximum room temperature of 30 °C or 35 °C depending on configuration and fan type, or not supported (N/A) in some configurations, and Dell's general notes say the most critical component sets a configuration's ambient. In the notes to its memory table for GPU configurations, Dell rates a 16 x 2.5-inch NVMe build with CPUs above 300 W TDP for 30 °C. For operation at ASHRAE A3/A4 ambients (40 °C and 45 °C), Dell requires its HPR SLVR fans in air-cooled systems, and in liquid-cooled systems with 2.5-inch drive configurations. That is one server model: read the matrix for yours.
4. 4.Run a sustained write test and read the SMART log before and after. Kioxia rates its CD8 family on the composite temperature the drive reports in SMART, with operating ranges that end between 72 °C and 76 °C across these pages; Kioxia's text does not tie each to a capacity. The NVMe definition says the platform should keep that temperature below the drive's warning threshold (wctemp), and that above the critical threshold (cctemp) the drive may throttle hard or shut down. We have not read a maker airflow (LFM) figure for any U.2 drive, so the test is the check.

   sudo nvme smart-log /dev/nvme0

   ReadingFor cooling, watch temperature against wctemp and whether 'Warning Temperature Time' rises across the test: a rise means the bay's airflow does not cover the drive's active power. The 'Thermal Management T1/T2' counters count only host-controlled thermal management and stay at 0 unless the host has set TMT1/TMT2, so a flat count proves nothing. The drop-out page linked below gives the full field readings.

Forum replies offer reasons, among them power-loss protection, higher IOPS and the supply rail the drive runs from; we have not read a maker document that sets out why, so treat them as opinions, as with the general forum warnings, carrying no figures, that U.2 drives draw a lot even at idle. What you can see for yourself is the power-state table. In one Level1Techs thread, the table posted for a consumer Samsung 980 Pro ended in two non-operational states, while the one posted for a Micron 9300 stepped from 25 W down to 10 W. In a Hacker News comment, an owner of Samsung 983 DCT drives (the retail PM983) reported no APST support and idle above 3 W for the M.2 version, which he called lower-power than the U.2 version. Linux APST can only use the non-operational kind.

Not according to the replies in a ServeTheHome thread on the P4510: they said a drive reaches its rated active figure under sustained heavy writing, not on short writes, and that writes draw much more than reads. If your drive reports active_power in nvme id-ctrl, the line after it names the workload the figure belongs to, '1MiB 32 RW, 30s idle' or '80K 128KiB SW'; that figure is the largest 10-second average under that workload.

Only if the link supports it, and LnkCap in lspci -vv says so for your drive. Threads say few U.2 drives support ASPM; in one, a reply named the Solidigm D7-P5520 as the only enterprise U.2 drive its author found that hints at ASPM support, and a later reply said Flexxon claims ASPM on a U.2 drive. We have not read a maker document stating which models support it, so check your own drive rather than a model list.

Not to find out. The wall test, the power-state table, APST and a set-feature cap cost nothing. A purchase is the answer only when the drives' rated active figures exceed your power or cooling budget, or when a drive's table offers no lower state and its idle draw matters more to you than its capacity. Choose a replacement from its maker's sheet: we found no maker document stating which models list non-operational states, and our drive list holds no power figure.