Do I need enterprise SSDs for Proxmox, or will consumer drives do?

1. 1.Get the host's writes per day from the drives you have now: the wearout page linked below measures them on NVMe (Data Units Written) and SATA (Total_LBAs_Written, or Logical Sectors Written where smartctl -l devstat is supported). Then turn each candidate's rating into writes per day and set the two side by side. A TBW rating spreads over its warranty: rated writes per day = TBW / (365 x warranty years). A DWPD rating means capacity x DWPD per day, on the lifetime the data sheet states.
   ReadingConsumer example: Samsung's 990 PRO data sheet gives 600TB (1TB) and 1,200TB (2TB) over 5 years, with coverage ending at the period or the TBW, whichever comes first: 600 / (365 x 5) is about 0.33TB a day and 1,200 / (365 x 5) about 0.66TB a day, roughly a third of each drive's capacity daily. Enterprise example: Samsung's PM893 page states its warranty as 5 years or 1.0 DWPD; by KIOXIA's definition one drive write per day is the full capacity written once a day, every day, for the stated lifetime. If your measured rate sits well under the consumer drive's figure, endurance alone does not decide the class.

1. 1.Run Proxmox's own benchmark against a path on the storage you want to test: replace <path> with a directory on that storage, for example /var/lib/vz. With no path, pveperf tests /. The result describes the device under the path you give, and nothing else.

   pveperf <path>

   ReadingRead FSYNCS/SECOND. The pveperf man page says the value should be greater than 200. On an otherwise idle host, a low figure points at how the storage under that path handles sync writes. Treat 200 as a floor rather than a target: in one TrueNAS forum thread, a consumer NVMe holding a Proxmox host's VM storage read about 755 while that host still showed high iowait.
2. 2.Identify the exact model and firmware, so you can find its data sheet and release notes.

   nvme id-ctrl /dev/nvme0 | grep -iE '^(mn|fr|hmpre|hmmin|vwc)\s'

   Readingmn is the model and fr the firmware revision. A non-zero hmpre means the controller supports Host Memory Buffer, which borrows host RAM; it does not prove the drive has no DRAM, so check the data sheet's DRAM line. vwc is printed as a hex value: bit 0 set (an odd value) means a volatile write cache is present. For a SATA or SAS SSD, smartctl -i /dev/sdX prints model, serial and firmware.
3. 3.To set a candidate drive beside Proxmox's published figures, run the paper's 4K sync-write test. WARNING: it destroys all data on the target device, and Proxmox's paper says so in its own words. Run it only on an empty drive, never on an rpool disk or any member of a live pool, and replace /dev/sdx with that drive.
   Destructive. This command destroys what is on the device and cannot be undone. Confirm the device path twice and have a backup of anything you still need before running it.

   fio --ioengine=libaio --filename=/dev/sdx --direct=1 --sync=1 --rw=write --bs=4K --numjobs=1 --iodepth=1 --runtime=60 --time_based --name=fio

   ReadingNote write IOPS and mean completion latency. On Proxmox's December 2020 testbed a Micron 9300 MAX 3.2TB reached 51,000 IO/s at 0.02 ms and an Intel DC S3500 120GB 12,099 IO/s at 0.08 ms, against 339 IO/s at 2.94 ms for a Samsung SSD 850 EVO 1TB, 254 IO/s at 3.93 ms for a Crucial MX100 512GB and 117 IO/s for a Seagate Constellation 7200.2 hard drive. Those numbers belong to those models on that testbed only. The paper recommends this test for every SSD before it goes into ZFS.

1. 1.Find the power-loss protection line for the exact model. Makers word it differently: KIOXIA lists 'Power loss protection (PLP)', Solidigm calls its scheme 'power-loss imminent (PLI) protection' with built-in self-testing, and Samsung's PM893 page says it 'prevents data corruption in case of power failure with power loss protection'. Our catalogue holds no PLP field, so this check is per model, in the maker's own document.
   ReadingWhy it matters: Samsung's 2016 white paper says SSDs usually acknowledge a write once it reaches volatile DRAM, that client SSDs do not normally carry capacitor-backed full PLP, and that its enterprise SSDs use capacitors to protect every write already acknowledged to the host. The 990 PRO data sheet we read lists no PLP.
2. 2.Read the endurance figure together with its basis. A DWPD figure means little without the lifetime and workload it is stated over: KIOXIA's CD8P-R page gives 1 DWPD with a 5-year warranty, while the Solidigm D7-P5510 brief we read states write endurance of up to 1 DWPD with no basis on that page.
   ReadingJEDEC's JESD218 sets separate Client and Enterprise application classes with their own requirements, and defines TBW as the terabytes a host writes to the SSD. We did not read the class requirements themselves (the standard text is paywalled), so this page states none of them.
3. 3.Look for a write figure stated for after the cache runs out. Samsung's 870 QVO data sheet rates sequential writes at up to 530 MB/s within Intelligent TurboWrite and 80 MB/s (1TB) or 160 MB/s (2-8TB) after it: once a write outlasts the cache, the post-cache figure applies. The 990 PRO sheet says its write figures were measured with TurboWrite active and gives no post-cache figure. In one Level1Techs thread, a poster running a Proxmox cluster on Crucial BX500s reported serious I/O bottlenecks during VM moves and restores; that is a forum report, not a documented cache effect.
   ReadingEnterprise pages describe their write behaviour as 'consistent performance' (KIOXIA CD8P-R) and 'predictably fast high performance' (Solidigm D7-P5510); those are the makers' claims, with no steady-state method on the pages we read. Seagate's over-provisioning blog says over-provisioning mitigates write amplification and, as general guidance rather than a per-drive figure, puts the space set aside at 7% to 28% of a drive's capacity, usually around 7-10% by default on consumer drives, and suggests considering 20-28% for professional and enterprise SSDs, especially in high-write environments.
4. 4.Compare the firmware revision found in the previous path with the maker's release notes before writing a consumer drive off. Samsung's notes for the 990 PRO include a September 2025 release 'To address the intermittent non-recognition and blue screen issue'.
   ReadingA known fix in the notes points at a firmware problem rather than the drive class. For drives that drop off the bus, follow the NVMe drop-out page linked below.

1. 1.List what backs the boot pool by full device path, then run the same for each VM pool.

   zpool status -P rpool

   ReadingEach path is a physical drive, and so a slot or bay a replacement must fit. Log or cache vdevs, if the pool has any, appear in the same listing.
2. 2.Boot / rpool. The guide's OS-storage line reads: hardware RAID with a battery-protected write cache, or non-RAID ZFS with an optional SSD for the ZIL. It names no SSD class for the OS disk. The same guide states that neither ZFS nor Ceph is compatible with a hardware RAID controller, so on a ZFS rpool each SSD's own write behaviour reaches the pool. If rpool also carries VM disks, the VM-storage wording applies to it as well.
   ReadingFor a host with only an M.2 2280 slot, Micron markets a 7450 model in M.2 2280 with power-loss protection, specifically for server boot use.
3. 3.VM storage. This is where the guide's 'Using consumer SSDs is discouraged' sits, beside 'SSDs with Power-Loss-Protection (PLP) are recommended for good performance'. The word is discouraged, not unsupported. The 2020 ZFS paper, a separate document, says M.2 NVMe drives can be used but that U.2 NVMe drives gave better disk performance than M.2 variants.
4. 4.ZFS log or cache device. Proxmox's ZFS chapter says a dedicated cache or log disk should be an enterprise-class SSD, and for log devices it recommends fast SSDs with power-loss protection because they have much smaller commit latencies. Of the three roles, this one carries the most specific wording in the current documents.
5. 5.Ceph, if it comes later. The OSD case has its own page with the Ceph-specific guidance and a latency diagnosis; a drive bought now can go through the sync-write test in the second path before it becomes an OSD, as Proxmox's ZFS paper recommends for every SSD before use.

It is the answer in the hardware FAQ of Proxmox's ZFS benchmark paper (testbed December 2020), which cites performance, reliability and endurance and calls enterprise-class SSDs with high endurance and power-loss protection essential for ZFS. The current admin guide's system requirements use a different word: consumer SSDs are 'discouraged', with PLP SSDs recommended for good performance. Quote each document with its own wording and date; one does not stand in for the other.

Check the physical fit first. In a Level1Techs thread about a Minisforum MS-A2, replies said 15mm U.2 and U.3 drives suit such boxes poorly for space and cooling, and that enterprise M.2 22110 drives fit some mini-PC slots. Enterprise models do come in M.2: Micron offers a 7450 in M.2 2280 with PLP for server boot, and Samsung offered its PM9A3 in E1.S, U.2 and M.2. An adapter is the other route.

A non-zero hmpre in nvme id-ctrl shows Host Memory Buffer support; it does not prove the drive lacks DRAM, which the data sheet's DRAM line settles. None of the primary documents behind this page state a performance cost for DRAM-less designs, so this page does not rank on it. Forum threads raise it: a ZFS pool of four WD Black SN770s with 200-300 ms write latency spikes ended unresolved, and replies to a PNY CS900 rpool with heavy IO delay pointed at DRAM-less, SLC-cached drives.

For the OS disk, the guide names hardware RAID with a battery-protected write cache as one option, and the pveperf man page mentions enabling write-back cache on a RAID controller with BBWC. Proxmox also states that neither ZFS nor Ceph is compatible with a hardware RAID controller, so once rpool or VM storage is ZFS or Ceph, the SSDs sit directly under the pool and their own sync-write behaviour is what counts.