How much RAM for TrueNAS, Proxmox, Unraid or a Synology

A NAS that only stores and shares files is well served by 8 GB on TrueNAS or Unraid and, on a Synology, by the 2 to 8 GB it ships with; a few apps take it to 16 GB, and each virtual machine adds its full memory on top. Proxmox VE is a hypervisor first: 2 GB for itself, the sum of what its guests are given, and a ZFS cache that a fresh install caps at 10 percent of RAM. The rule most forums still quote, 1 GB of RAM per TB of storage, traces to the FreeNAS 8 documentation wiki, archived in 2012, and appears in no current TrueNAS or OpenZFS document; iXsystems’ own TrueNAS Mini R ships with 32 GB for up to 216 TB of raw storage.

The whole answer, by platform

The vendor figures carry their version; the three middle columns are this guide’s tiers. Each row links the section that argues for it.

The tiers rest on the vendors’ figures, the 16 and 32 GB in iXsystems’ TrueNAS Minis and the apps’ own figures. iXsystems’ 16 and 32 GB for VMs describe a TrueNAS box serving iSCSI storage to VMs on other hosts; a VM running on TrueNAS itself is counted at its own size. What decides the purchase is a sum - the platform’s base, every VM at its full size, each app, a cache budget if the pool is ZFS - and two limits above it: the box’s maximum, which on a Synology is 6, 8 or 32 GB by model (the RS422+ is fixed at 2 GB), and the price of the module type it takes. The column you need is set by the virtual machines you run, not by the size of the pool.

The checklist

1. Write the sum before you open a listing: base, every VM at full size, each app, a cache budget, headroom (the terms, four builds).
2. Find the box’s ceiling first. It is the box maker’s figure: HPE caps the MicroServer Gen10 Plus at 32 GB though its Xeon E-2224 addresses 128 GB (the tables).
3. Match the module type: SO-DIMM or full-size, unbuffered or registered, ECC or not; a Ryzen Synology takes ECC SO-DIMMs. Registered vs unbuffered has the rules.
4. Count every VM at its full allocation. Ballooning guarantees only a minimum, and a passthrough VM cannot balloon on Proxmox or Unraid.
5. Count nothing for swap or KSM. TrueNAS removed swap management in 24.10, and Proxmox’s KSM starts merging only when memory is nearly committed.
6. Set the ZFS cache on purpose; each platform’s default is in the ARC table.
7. Leave deduplication off unless zdb -S simulates a ratio comfortably above 2x (deduplication).
8. Buy RAM before an L2ARC. Each block it caches costs 96 bytes of ARC.
9. Fill a Synology in matched pairs, with the warranty clause in view (upgrading a Synology).
10. Test new memory, then read “used” correctly: a full graph on a NAS is cache (how to see it).

How the number adds up

Every platform here is sized by the same five terms, and vendors publish only the first, which is why their minimums look small. Desktops, workstations and local AI are sized in how much RAM do you need.

The platform base is small

TrueNAS’s 8 GB covers the system and ZFS for up to eight drives; Proxmox VE asks 2 GB; Unraid publishes no figure; Synology’s DSM keeps 1.5 GB of an 8 GB unit when it hosts VMs, and 10 percent from 16 GB up. These are floors, not purchase figures.

Virtual machines are counted whole

TrueNAS’s 25.10 documentation: “While a VM is running, these resources are not available to the host computer or other VMs.” Proxmox VE guarantees a ballooned VM only its minimum, so a host sized to the minimums makes guests swap when they reach for the rest, and a VM with a PCIe or GPU passthrough device cannot balloon at all. Overhead comes on top. Proxmox’s 9.0 release notes say VMs “usually consume a higher amount of memory on the host than the amount that is reported from inside the guest”, with no figure; Synology publishes one, 128 MB per VM plus 80 MB per virtual CPU and 32 MB per virtual disk.

Apps and containers are limits, not reservations

TrueNAS’s Apps catalog gives apps such as Plex, Jellyfin and Immich a 4096 MB memory limit by default; Docker Engine and Unraid’s Docker set none; a Proxmox LXC container defaults to 512 MiB. A cap overstates what a quiet machine uses; with no cap, Docker warns, a host that runs out lets the kernel’s out-of-memory killer take any process.

The cache takes what is left

On ZFS the read cache is the ARC, and each platform sets how much of the remainder it may take (the ARC table). XFS and Btrfs disks in an Unraid array, and a Synology volume, use the ordinary page cache, the subject of RAM caching.

Headroom, because swap will not save you

TrueNAS 24.10 removed swap management, and a TrueNAS staff member wrote on May 4, 2024: “We do not recommend using a SWAP drive. ZFS expects ARC to be in RAM.” Proxmox warns that swap on a ZFS zvol can block the server and advises leaving 1 GB free for the host. Treat RAM as a hard budget and leave room above the sum.

TrueNAS: 8 GB for eight drives, then the workload

As of October 15, 2026, iXsystems recommends TrueNAS Community Edition 25.10.7, from September 2, 2026, for general users. TrueNAS 27, renamed from 26, is at its first release candidate (October 6, 2026), for early adopters, with OpenZFS 2.4. TrueNAS CORE ended with 13.3-U1.2 on April 29, 2025, and iXsystems names 25.10 as its migration path.

The sizing rule in the hardware guide

The 25.10 hardware guide’s minimum is a 2-core x86-64 processor, 8 GB of memory and a 16 GB boot SSD; the installation page says “TrueNAS requires at least 8GB of RAM (more is better) and a 20GB Boot Device.” The TrueNAS 27 guide keeps 8 GB with “at least” added. The rule that scales it has been the same from 22.12 through 27 and in CORE 13: “You should have 8 GB of RAM for basic TrueNAS operations with up to eight drives. … Add 1 GB for each drive added after eight to benefit most use cases.” It counts drives, not terabytes: a 12-drive box starts at 12 GB, a 24-bay chassis at 24 GB, before any app or VM. It first appeared in the FreeNAS 11.3 guide.

Then the guide names what actually moves the figure: “A 20 TB pool backing many high-performance VMs over iSCSI might need more RAM than a 200 TB pool storing archival data.” For a box that backs VMs on other hosts over iSCSI it asks for at least 16 GB, with 32 GB or more for optimal performance, and it adds 2 GB for Active Directory through Winbind, more for each app and VM it runs itself, 5 GB per TB of deduplicated storage and about 1 GB per 50 GB of L2ARC.

Apps, containers and virtual machines

TrueNAS 24.10 moved Apps from Kubernetes to Docker, claiming faster deployment and lower idle CPU but nothing about memory. The catalog’s Plex, Jellyfin and Immich apps default to a 4096 MB memory limit. One mismatch is worth fixing: Immich asks for at least 6 GB, and its FAQ warns that “memory constraints work by terminating the container”, so if Immich’s machine-learning jobs die, raise its limit first.

The 25.10 VM tutorial sets a 256 MiB minimum per VM, offers minimum and maximum memory for ballooning, and warns that a VM given all available memory “can slow the host system or prevent other VMs from starting”. An iXsystems engineer wrote in August 2022 that the middleware lowered the ARC’s maximum when VMs were configured, and 25.04 and 25.10 list fixes for the ARC not giving memory back fast enough to VMs (NAS-135904, NAS-135499): run a current release on a box that mixes a large cache with VMs. LXC containers, new in 25.04, can run without a limit, and “Memory allocated to containers reduces available memory for TrueNAS ZFS caching.”

The cache took half before 24.04 and nearly all of it since

Until TrueNAS SCALE 24.04 the ARC stopped at half of RAM, the OpenZFS Linux default. The 24.04 “Dragonfish” notes record the change: “ZFS ARC memory allocations are updated and behave identically to TrueNAS CORE.” With kernel 6.6’s Multi-Gen LRU code, the larger ARC caused swap thrash on 24.04.0, and 24.04.1 (May 28, 2024) disabled swap and Multi-Gen LRU. TrueNAS 25.04 moved to OpenZFS 2.3, whose ceiling is the larger of memory minus 1 GiB and five-eighths of it. On a 64 GB TrueNAS box the ARC may grow to about 63 GiB, where SCALE before 24.04 stopped it at 32 GiB, so extra RAM is no longer half wasted, and advice from 2023 or earlier to raise zfs_arc_max by hand is out of date.

What iXsystems puts in its own boxes

Every current TrueNAS Mini ships with ECC DDR4: the Mini X with 16 GB (to 32 GB, for up to 90 TB raw), the Mini X+ with 32 GB (to 64 GB, 105 TB) and the Mini R with 32 GB (to 64 GB, 216 TB); the Mini X fills its four slots in same-capacity pairs. The R20 manual gives 32 to 192 GB against up to 2.9 PB raw with expansion: by this guide’s arithmetic about 0.15 GB per raw TB in the Mini R and 0.07 in a full R20. iXsystems does not follow the per-terabyte rule in its own hardware. The disks are the drive site’s drives for a NAS guide.

Proxmox VE: 2 GB, every guest, and a cache you size yourself

Proxmox VE 9.2 was released on May 21, 2026, and this section follows its documentation version 9.2.13 of September 24, 2026; Proxmox VE 8 reached end of life in August 2026. The installation chapter’s minimum, 1 GB plus guest RAM, is labeled for evaluation only. The recommendation is “Minimum 2 GB for the OS and Proxmox VE services, plus designated memory for guests. For Ceph and ZFS, additional memory is required; approximately 1GB of memory for every TB of used storage.”

Four formulas from one vendor

Proxmox is the one vendor here whose general rule is still per terabyte, and its documents give four formulas, to be quoted rather than merged:

For a mirror of two 8 TB drives holding 5 TB, the four give: 7 GB (the 2 GB base plus 5 GB for the used terabytes), 20 GB (4 GB plus 16 TB raw), an ARC of 9.3 GiB (2 GiB plus 7.3 TiB available) and at least 8 GB. The ARC rule is the one to buy with, because it budgets one line of the sum rather than the whole machine.

The 10 percent cap

Since Proxmox VE 8.1 (November 23, 2023) the installer writes a limit into /etc/modprobe.d/zfs.conf: “ZFS uses 10 % of the host memory, clamped to a maximum of 16 GiB, for the Adaptive Replacement Cache (ARC) by default.” Version 8.4 added a clamp that keeps at least 1 GiB free on small systems.

On a small host the default sits well below Proxmox’s own rule, so a 32 or 64 GiB box with a multi-terabyte pool should raise it (the Proxmox build shows how). A host installed before 8.1 keeps the ZFS default, which Proxmox gives as 62.5 percent on ZFS 2.3 and later and the OpenZFS manual as memory minus 1 GiB. Above 256 GiB, a low zfs_arc_max is ignored unless zfs_arc_min, 1/32 of memory by default, is lowered below it.

Ballooning, KSM and containers

Ballooning hands extra memory to a VM only while host usage is under a target, 80 percent by default; under pressure the guest swaps and, in the last resort, runs its own out-of-memory killer. Proxmox does not recommend the Windows balloon driver on critical systems. Kernel Samepage Merging is on by default, Proxmox warns it exposes VMs to side-channel attacks, and by a reading of Proxmox’s ksmtuned script it only starts merging near 80 percent of RAM committed. Buy for each VM’s full allocation. Ballooning guarantees only the minimum, and a passthrough VM always holds all of its memory.

An LXC container shares the host kernel, its memory a cgroup limit; Proxmox calls containers lightweight but publishes no saving. One app maker steers the other way: Immich “runs well” in “a full virtual machine” and says Docker in LXC “is not recommended”, so budget Immich on Proxmox as a VM.

Ceph

The 9.2 Ceph chapter recommends at least 8 GiB per OSD, the daemon’s default being 4 GiB, and warns that recovery and backfill use more. Upstream Ceph’s Tentacle documentation, the release Proxmox VE 9.2 ships, sizes a node at more than OSDs times osd_memory_target times two, plus 20 percent for spikes. A three-node homelab with four OSDs per node needs at least 32 GiB per node by Proxmox’s figure, and about 38 GiB by upstream Ceph’s (4 x 4 GiB x 2, plus 20 percent), before any guest.

Proxmox Backup Server

Proxmox Backup Server 4.2 (documentation of October 2, 2026) asks for “minimum 4 GiB for the OS, filesystem cache and Proxmox Backup Server daemons. Add at least another GiB per TiB storage space”, so about 24 GiB for a 20 TiB datastore. For hard-drive datastores it calls a special device mirror for metadata “highly recommended”. A Backup Server installed on ZFS gets an ARC cap of 50 percent from its installer, not 10 percent.

Unraid: an OS that lives in RAM and a 20 percent ZFS cache

Unraid 7.3.3 (October 8, 2026, with ZFS 2.4.4) is the current stable release. It boots from a USB flash drive into “a root RAM file system”, so the whole operating system lives in memory.

There is no published minimum

Until at least August 2025, Unraid’s documentation said “A minimum of 4 GB of RAM for basic NAS functionality”. That page now redirects, and no current page states an OS minimum; the “4 GB” on unraid.net’s getting-started list is the size of the USB drive. The only current figures are for VMs: 8 GB minimum, “16 GB or more (add RAM for each active VM)”. Unraid 7.0 noted that memory pressure can push system files out of RAM to be re-read from flash; a fastusr flag keeps them resident for about 500 MB. Size an Unraid file server at 8 GB and treat 4 GB as history; the 8 GB is this guide’s whole-machine figure, borrowed from Unraid’s own VM minimum.

XFS and Btrfs disks use the page cache; ZFS uses the ARC

An array of XFS or Btrfs disks uses the Linux page cache; a ZFS pool, or an array disk formatted ZFS (allowed since 6.12), uses the ARC and the same cap. Unraid calls the per-terabyte rule “outdated advice”. From 6.12.0 the ARC was limited to one eighth of memory; from the 7.3.0 tag (May 12, 2026) Lime Technology’s boot script sets 20 percent instead, with no release note, and the ZFS documentation page still says “usually 1/8th”. Since 7.3.0 the cap has a control at Settings, Disk Settings, Tunable (zfs_arc_max): presets from 10 to 90 percent, a custom figure, or “Unlimited (Dynamic)”, which writes 0 and hands the ceiling back to OpenZFS.

By a reading of the 7.3 boot script, once a zfs_arc_max line exists on the flash drive the 20 percent is not recomputed, so after adding RAM pick a preset again and check /sys/module/zfs/parameters/zfs_arc_max.

Docker, virtual machines and buying memory

Unraid’s Docker containers have no memory limit by default; 7.4.0-rc.1 adds an optional one. For VMs, Unraid suggests 1 to 2 GB for a virtual server, 4 to 8 GB for a virtual desktop such as Windows 11 and 8 to 16 GB or more for a GPU passthrough gaming VM, adding that memory is “only consumed while VMs are running”. Its stability page says to buy from the motherboard’s qualified list, run at SPD speed because “Intel XMP and AMD AMP profiles are overclocks”, do not mix brands or speeds, and run Memtest86+ for two to four hours. The documentation never mentions ECC.

Synology: a ceiling per model and a warranty clause

A Synology is the easiest box here to size and the hardest to exceed: each model has a published maximum. The current DSM is 7.4.1-90080 Update 1 of October 12, 2026.

Plus, neo+ and play models: installed, maximum, slots and type

On the one-slot Celeron models the factory memory is not in the slot: the installation guides describe “an empty memory slot”, so the upgrade adds a 4 GB module. The RS422+ lists no slot, maximum or memory accessory, and Value and J-series units such as the DS223, DS423, DS124 and DS223j list no slot either: buy them at the size you need. DSM 7.4 is the last version for the DS918+ and DS418play.

Synology’s own parts are D4NS01-4G for the current Celeron units and the neo+ variants, D4NESO-2666-4G on older Celeron sheets, D4ES01 to D4ES04 for the ECC units and D3NS1866L-4G for DDR3L. Third-party equivalents are the same physical types: DDR4 SO-DIMMs, sorted by the ECC column for the ECC units, and DDR3 SO-DIMMs. The ECC units take ECC SO-DIMMs: the small 260-pin laptop size, unbuffered, with a ninth chip per rank. A full-size ECC UDIMM will not go in, and a listing that says “ECC” without “SO-DIMM” or “260-pin” is usually the wrong part. Speed buys nothing: the specifications say modules “operate at the maximum frequency specified by the CPU manufacturer”, and Intel and AMD give DDR4-2400 for the four Celeron and Ryzen DDR4 processors, so a faster module runs at 2400 MT/s or less.

Upgrading: Synology modules, third-party modules and the warning

Every current Plus specification says: “Synology will not provide complete product warranty or technical support if non-Synology memory modules are used for memory expansion.” A Synology white paper of July 2026 claims third-party memory drew more than nine times as many stability tickets, without publishing a method. No Synology document says DSM refuses third-party memory on a Plus model: the white paper describes a startup check only on “enterprise-tier platforms”, and community reports describe a “Detected non-Synology memory module” notification while the memory keeps working. The cost of a third-party module on a Plus model is the warranty and support clause, not a refusal to boot.

The installation rules are stricter. On the two-slot models “all modules must be identical and share the same model number”, and the installation guides list only matched pairs: 4 + 4, 8 + 8, 16 + 16. The DS723+ table goes from its single 2 GB module straight to pairs, so upgrading means replacing it, while the DS918+ and DS418play guides say “Do not remove the pre-installed memory module to prevent voiding product warranty”. Mixing ECC and non-ECC modules switches ECC off. Afterward, run Synology’s memory test from Synology Assistant three times, about 1 to 1.5 hours each at 16 GB or less.

What DSM features need memory

DSM has no per-terabyte rule; its memory floors are attached to features.

Published capacity figures assume a full unit: the current specifications say models with expandable memory “are tested with the maximum supported quantity installed”, and the DS920+ sheet gives 1,000 concurrent SMB, AFP and FTP connections, 2,000 “with RAM expansion”.

ZFS on any platform: the ARC, deduplication and L2ARC

Every ZFS platform here runs OpenZFS, current release 2.4.4 of August 21, 2026; each platform sets its own defaults.

The ARC ceiling, release by release

With zfs_arc_max at 0, OpenZFS 2.3 and 2.4 cap the ARC at the larger of memory minus 1 GiB and five-eighths of memory, which above about 2.7 GiB is memory minus 1 GiB. The 2.2 branch, still maintained, keeps the old Linux default of half.

The ceiling is not a reservation: the ARC gives memory back under pressure down to a floor of 1/32 of memory, and lowering zfs_arc_max on a running system does not shrink it until there is pressure (RAM caching has the mechanics). A large ARC is not a leak; an ARC capped far below the working set is a choice that costs disk reads. One OpenZFS page has not caught up: Workload Tuning still says “half of system memory on Linux”.

Deduplication is the one per-terabyte rule

Deduplication keeps a table of every unique block, and the table must sit in RAM to be fast. The published figures disagree:

Block size, not capacity, sets the bill; RAM caching works the arithmetic. OpenZFS warns that an undersized system can end up with a pool it cannot import. Fast dedup, enabled in TrueNAS 25.04, adds a dedup_table_quota cap; iXsystems claimed “up to 90%” less memory, an OpenZFS collaborator said in December 2024 it would “reduce a bit, but not dramatically to change recommendations”, iXsystems supports it only on Enterprise systems with 256 or 512 GB, and Unraid calls dedup “risky for most users”. Simulate with zdb -S first; OpenZFS says if the ratio “is not comfortably above 2x, dedup is not worth it.”

L2ARC costs RAM, and a special vdev saves none

Every block an L2ARC holds keeps a 96-byte header in the ARC, and OpenZFS says “L2ARC costs RAM. … On a RAM-constrained system an L2ARC can make things slower.” On a default Proxmox install of a 32 GiB host, with a 3.2 GiB ARC, the default 33 percent header cap leaves roughly 1 GiB to index an L2ARC, by this guide’s arithmetic. Buy RAM first; a special vdev is storage rather than cache and lowers no memory figure, and sizing either flash tier is the drive site’s NVMe caching guide.

ECC on a NAS: what each vendor says

No NAS operating system in this guide requires ECC memory, and each ZFS vendor’s documents lean toward it. What the extra chip corrects, which boards switch it on and how to prove it is running are in ECC vs non-ECC. The positions, as of October 15, 2026:

• TrueNAS: “most users strongly recommend ECC RAM as another data integrity defense”, but “Many TrueNAS systems operate every day without ECC RAM”, and “RAM failures usually occur in the first three months, so test all RAM before deployment.”
• OpenZFS: bit flips “can have fairly dramatic consequences for all computer filesystems and ZFS is no exception”; the FAQ says it “isn’t really a requirement, but it’s highly recommended”.
• Proxmox: “we recommend the use of high quality ECC RAM”; a Proxmox staff member wrote in 2020 that “The rumor that ZFS really needs ECC is just that. A rumor.”
• Unraid: not mentioned.
• Synology: decided by the model; the Ryzen Plus units ship ECC, except the neo+ variants (announced August 2026), which ship 4 GB non-ECC and list both Synology’s non-ECC and ECC modules; the Celeron processors cannot use ECC.

For data you cannot re-create, ECC is worth it where the platform makes it cheap, and the platform decides that before the module does. On a non-ECC build, test the memory long; an iXsystems engineer’s request to a forum user with desktop hardware and pool metadata corruption, in April 2024, was “please run good long memory test”.

What the apps ask for

The last column matters: a whole-machine figure includes an operating system the NAS has already paid for, so adding them up over-counts.

Machine learning is what swings these numbers: Immich drops from 6 GB to 4 GB without it, Frigate doubles with its enrichments, and each extra Immich machine-learning worker loads its own copy of the models. Frigate keeps frames in shared memory, about 210 MB for two cameras at 720p against Docker’s 64 MB default, which a Frigate container needs raised. Transcoding Plex or Jellyfin in /dev/shm is a community practice, described in an Unraid forum guide, not something either project documents.

Four builds added up line by line

Each build adds the vendor figures above; this guide’s assumptions are labeled. Prices are median eBay US Buy It Now asking prices, including shipping, read on October 15, 2026, not sale prices (what each tier costs).

A four-bay TrueNAS box: 4 x 12 TB, Jellyfin and Immich

An HPE MicroServer Gen10 Plus with a Xeon E-2224, four 12 TB drives and TrueNAS 25.10:

TrueNAS 25.10, up to eight drives (hardware guide)            8 GB
Jellyfin, at the catalog's default memory limit               4 GB
Immich, limit raised to Immich's own minimum                  6 GB
Committed                                                    18 GB

Installed: 2 x 16 GB DDR4 ECC UDIMM, the machine's maximum   32 GB
Left for the ARC and headroom                                14 GB
Two modules at the median asking price, $210 each           $420

For comparison: "1 GB per TB" on 48 TB raw                   48 GB
Proxmox's "4 GB plus 1 GB per TB raw" on the same disks      52 GB

Committed is the sum of caps: Jellyfin and Immich at rest use well under their limits, which is why the 16 GB tier holds for a few apps, but a 16 GB box would leave the ARC almost nothing once Immich’s machine learning runs. The folklore figure is more than this machine can hold, while iXsystems asks for 8 GB. HPE’s QuickSpecs (August 18, 2025) allow only ECC UDIMMs, 2 x 16 GB, at 2666 MT/s with this processor: DDR4 ECC filtered to unbuffered parts at 16GB. If the pool later serves VMs on other hosts over iSCSI, the hardware guide’s optimal figure, 32 GB or more, is this machine’s ceiling, and the Gen10 Plus v2 at 64 GB is the next step.

A Proxmox host: Home Assistant, a Windows VM, three containers and a ZFS mirror

An HPE MicroServer Gen10 Plus v2 with a Xeon E-2314, Proxmox VE 9.2 and a mirror of two 8 TB drives, about 7.3 TiB usable:

Proxmox VE, OS and services (installation chapter)            2.0 GB
Left free for the host (qm chapter rule of thumb)             1.0 GB
ZFS ARC by Proxmox's rule: 2 GiB + 1 GiB x 7.3 TiB            9.3 GiB
Home Assistant OS VM (HA's example value)                     4.0 GB
Windows 11 VM (4 GB minimum; 8 GB assumed for desktop use)    8.0 GB
LXC: Pi-hole                                                  0.5 GB
LXC: Jellyfin (4 GB "may be enough" on headless Linux)        4.0 GB
LXC: Nextcloud (the tuning guide's 2 GB PHP example)          2.0 GB
Total                                                         about 31 GB

Installed: 2 x 32 GB DDR4 ECC UDIMM, the machine's maximum    64 GB
Two modules at the median asking price, $396 each            $792

Thirty-one gigabytes would fill a 32 GB machine on day one, with nothing for VM overhead; 64 GB leaves room for Immich as its own VM at its recommended 8 GB. One setting needs changing: the installer’s cap on a 64 GB host is 6.4 GiB, below the 9.3 GiB Proxmox’s own rule asks for. The installer’s line in /etc/modprobe.d/zfs.conf becomes:

options zfs zfs_arc_max=10737418240

Then, as root:

echo 10737418240 > /sys/module/zfs/parameters/zfs_arc_max   # applies now
update-initramfs -u -k all    # root on ZFS: makes the limit hold at boot

The second command matters on the install that gets the 10 percent cap, a root on ZFS, whose module loads from the initramfs. Run the Windows VM without the balloon driver and budget its full 8 GB. HPE lists the v2’s 32 GB modules as 2Rx8 ECC UDIMMs at 3200 MT/s and does not recommend mixing capacities; they are on 32GB DDR4 filtered to unbuffered ECC.

An Unraid array with Docker apps

A desktop board with two DDR4 slots, Unraid 7.3, an XFS parity array and an SSD cache pool, running Plex, Immich, Paperless-ngx and Pi-hole in Docker and a Windows 11 VM some evenings:

Unraid itself, this guide's file-server figure                8.0 GB
Plex (Plex's figure for a whole server)                       4.0 GB
Immich with machine learning (Immich's minimum)               6.0 GB
Paperless-ngx, one web worker (no figure published; assumed)  1.0 GB
Pi-hole                                                       0.5 GB
Windows 11 VM while running (Unraid: 4 to 8 GB)               8.0 GB
Total                                                        27.5 GB

Installed: 2 x 16 GB DDR4                                     32 GB
Two desktop non-ECC modules at the median, $153 each         $306

This sum over-counts on purpose: Plex’s and Immich’s figures each include an operating system Unraid already has, and the surplus is the headroom that no swap provides. Without the VM it is 19.5 GB, past 16 GB, so 32 GB is the step either way. The parity array uses the page cache, so the spare memory caches reads with no setting; a ZFS pool on 7.3 would take up to 6.4 GiB of the 32. On a Ryzen board whose maker lists ECC support, two unbuffered ECC 16 GB modules were $420 at the median.

A Synology DS923+, from 4 GB to 32 GB

The DS923+ ships with one 4 GB ECC SO-DIMM in one of two slots and takes 32 GB as 16 GB x 2. Matched pairs only, so the path has three steps:

Factory, 4 GB              Photos object recognition: yes
Add a 4 GB: 4 + 4 = 8 GB   DSM keeps 1.5 GB; 6.5 GB for VMs
Replace: 8 + 8 = 16 GB     DSM keeps 10 percent; 14.4 GB for VMs
Replace: 16 + 16 = 32 GB   28.8 GB for VMs; Active Backup at 50 jobs

Two 8 GB DDR4 ECC SO-DIMMs at the median, $130 each        $260
Two 16 GB DDR4 ECC SO-DIMMs at the median, $225 each       $450

The second 4 GB must share the factory module’s model number, D4ES02-4G on the DS923+ sheet, so in practice it is Synology’s own; read the installed module’s label first. Two small VMs show the reservation: Home Assistant OS at 2 GB and Ubuntu Server at 3 GB, each with two vCPUs and one virtual disk, cost 5 GB plus 2 x (128 + 160 + 32) MB of overhead, about 5.6 GB, which fits in the 6.5 GB an 8 GB unit leaves. Synology recommends four 1 GB VMs on a DS923+ even at 32 GB, half the V1500B units’ eight, which suggests the dual-core R1600 rather than memory sets that limit. A third-party DDR4 ECC SO-DIMM of 16 GB or less works, at the cost of the warranty clause. Synology itself fits a non-ECC D4NS01-4G module in the V1500B neo+ variants, but mixing ECC and non-ECC modules switches ECC off.

Which processors and machines take which memory

The processor sets the outer bound and the box maker sets the real one. Intel lists the Xeon E-2224 at 128 GB; HPE’s MicroServer Gen10 Plus built on it stops at 32 GB. Intel’s own page for the Xeon E-2124 says its 128 GB figure “requires both a BIOS update” and “hardware platform support”. The other limits that can bind are in server memory capacity limits.

Processors

The N-series parts are 16 GB, one channel and non-ECC by Intel’s specification; its datasheet (January 2025) says their In-Band ECC is “Supported only in Chrome systems”, and reports of 32 or 48 GB working are community tests, one later reported unstable under load. On Supermicro’s Atom C3000 and Xeon D boards, the maximum needs registered memory: the A2SDi-8C+-HLN4F takes 256 GB of RDIMM but 64 GB of UDIMM, and the X10SDV-4C-TLN2F, with a Xeon D-1521, 128 GB against 64 GB. 12th- to 14th-generation Intel Core chips do ECC only on a W680 board, because ECC “requires both processor and chipset support”; Intel’s workstation guide lists the i5-14500 among them. AMD’s “Yes” depends on the board: the Ryzen 5 5600 page uses the same wording, and AMD names no boards. Filling four DDR5 slots on AM5 drops the rated speed to DDR5-3600, which matters when a Proxmox build wants 128 GB from four modules.

Machines

HPE’s QuickSpecs for both MicroServers rule out non-ECC UDIMMs, RDIMMs and LRDIMMs, and with them the cheapest used server memory. On the T140 a Dell staff member wrote in February 2021 that “32GB DIMM is not supported and not validated”; a user’s 4 x 32 GB report in the same thread is a community result. Each machine’s modules, ranked by price per gigabyte, are on its page, and every server the site covers is on memory by server.

The rules that are folklore, and where each came from

Most confident NAS memory rules started as something true in one document at one date.

The common thread is a correct figure that outlived its context: a 2012 hedge for large FreeBSD pools, a 2022 recommendation withdrawn the same year, an upgrade its own guide called unofficial.

What each tier costs

Prices are median eBay US Buy It Now asking prices from the site’s listing data, read on October 15, 2026, including shipping and excluding listings sold for parts: asking prices, not sales, in a market that moves weekly. Each listing page has the live figures, trends has their history, and the methodology page explains how listings are recorded.

A tier’s cost is the median times the module count, an estimate rather than a quote. “Per GB” is the median of each listing’s own price per gigabyte, not the median price divided by capacity, so the two columns need not agree to the cent. Registered DDR4 is the cheapest modern gigabyte: a used rack server’s 256 GB came to $2,264 against $1,870 for 64 GB of DDR5 ECC, but registered modules fit only machines that take them, and a MicroServer is not one. DDR5 ECC costs two to three times DDR4 ECC per gigabyte, $29.21 against $12.37, and a 32 GB DDR5 RDIMM’s median was $1,296, about 4.6 times its DDR4 equivalent, so the platform is the largest cost decision in a homelab build (DDR4 vs DDR5). DDR3 servers are the cheapest memory of all, on platforms whose power draw is its own argument.

How to see what your system is really using

A NAS that has run for a while shows its memory as nearly full, because spare memory becomes cache.

On any ZFS system, read the ARC itself. On OpenZFS 2.4 and later the tools are zarcstat and zarcsummary; before 2.4.0 they were arcstat and arc_summary, the names on an OpenZFS 2.3 system such as TrueNAS 25.10.

zarcstat 5                                      # ARC size and target, every 5 s
zarcsummary | less                              # counters since boot
cat /sys/module/zfs/parameters/zfs_arc_max      # 0 means the OpenZFS default
grep l2_hdr_size /proc/spl/kstat/zfs/arcstats   # RAM held by L2ARC headers

Whether more RAM would help is told by the ghost-list hit rate, the reads a larger ARC would have caught; RAM caching shows how to read it. A flat ghost-hit line means more memory buys the cache nothing.

On TrueNAS 25.10, the dashboard’s memory card is a gauge “showing GiB Free in blue, ZFS Cache in fuchsia pink, and Services in orange”, so the cache stands apart from what the system and apps hold. On Proxmox VE 9.0 and later, the node’s memory graph shows the ARC as its own series, so a full bar is no longer mistaken for guests. On Unraid, the Dashboard shows ZFS and Docker memory, and the documentation says a low “available” reading “doesn’t necessarily indicate a problem” because Linux caches aggressively. On a Synology, DSM’s help says memory use “remains high because the system stores frequently accessed data in cache”, and the knowledge base says to add Buffer, Cached and Free in Resource Monitor to get the memory actually available. Real shortage shows as high swap use and, since DSM 7.0, a “Your system is out of memory” notification.

On any platform, test new memory before trusting it, and test it long. On an ECC machine, read the corrected-error counters after the test as well, because a test can pass while the controller corrects a failing module underneath it; ECC vs non-ECC has the procedure.

What to do with this on a listing page

1. Check the form factor against the box’s table. Every Synology here takes SO-DIMMs, as do most mini PCs; the MicroServers, towers and rack servers take full-size DIMMs (the machines).
2. Check ECC and registration as a pair. The MicroServers, ML30 and T150 want unbuffered ECC; a Ryzen Synology wants an unbuffered ECC SO-DIMM; only the rack servers take registered modules. A title that says “ECC” alone does not say which.
3. Check the capacity per module against the slot count. A two-slot box reaches 32 GB only with 16 GB modules and 64 GB only with 32 GB modules; a one-slot Celeron Synology takes one 4 GB module and no more.
4. Buy a Synology’s modules as an identical pair, the same model number in both slots; HPE advises against mixing capacities in the Gen10 Plus v2.
5. Pay nothing for speed above the platform’s rate: 2400 MT/s in the Celeron and Ryzen Synology units, 2666 MT/s in a Gen10 Plus with the Xeon E-2224, 3200 MT/s in a v2 with the E-2314.
6. Read the label in the photographs, not the title, for the part number, rank and ECC, then run the checks in buying used RAM on eBay.