A NAS build is really one decision wearing a lot of disguises: how many drives do you need to spin, now and in three years? Everything else, the board, the RAM, the power supply, the chassis, follows from the bay count. Get that number right and the rest of the parts list almost writes itself. Get it wrong and you either overpay for empty bays or outgrow the box in a year.
This guide lays out three tested ZFS storage server builds around that one question. A quiet four-bay box for backups and a media library, an eight-bay tower with ECC and 10GbE for a serious homelab, and a used-enterprise rackmount that drives 24 hot-swap bays off a single card. All three run TrueNAS SCALE on ZFS, all three use ECC memory, and none of them put an SMR drive anywhere near the pool. The one thing that has changed the math this year is memory, and now the drives too. Re-checking every line against live Amazon listings in September 2026, and again in October, moved almost all of them: a 32GB DDR5 ECC UDIMM module is about $850, so the tower build’s 64GB now costs more than its board, CPU, case and power supply put together. Spinning disks went the same way: against the prices we published for these builds in June 2026, a 4TB CMR NAS drive moved from $100-$120 to about $280 where new US stock exists, and a 20TB drive from $280-$400 to about $880 for the cheapest current listing. Stock at these capacities keeps flipping between the WD Red Plus and Red Pro lines: in September the Red Plus 4TB had thinned to resellers, in early October the Red Pro 4TB went out of stock, and by late October the Red Plus 4TB had almost no new stock besides a Japan-market import. The starter table now names the WD Red Plus 6TB, the smallest Red Plus with a normal new listing and still a quiet 5400 RPM drive, while at 8TB the Red Pro has no normal new listing either. That reprices a populated pool more than it reprices the machine. Two consequences worth acting on: buy memory and drives first and check both the day you order, and if you can start with fewer DIMMs, do, because both consumer boards here have four slots and ZFS will happily take more ARC later.
Current as of October 2026.
The three builds at a glance
The standout for most people is the eight-bay tower. It clears 64GB of ECC, runs an IT-mode HBA, and gets you onboard dual 10GbE without an add-in card, which is the sweet combination for a homelab that runs apps and VMs off the pool rather than just storing files. The four-bay starter is the right call if you want a quiet, low-power box for backups and Plex and you are learning ZFS on real hardware for the first time. The rackmount beast is for the person who already knows they want 100TB-plus and is comfortable buying used server gear off eBay.
- Quiet 4-bay starter (~$2,060 to $2,335 excl. drives): Ryzen 5 7600, ASRock Rack ECC board, 32GB ECC UDIMM, Jonsbo N6. Backups, media, learning ZFS.
- Eight-bay tower (~$3,260 to $3,485 excl. drives): Ryzen 7 7700, ASRock Rack board with dual 10GbE, 64GB ECC UDIMM, 9300-8i HBA, SilverStone CS382. The all-rounder.
- Rackmount beast (new parts only come to $320 to $350; the platform itself is a used-market buy): used EPYC 7302, ASRock Rack ROMED8-2T, 128GB ECC RDIMM, 24-bay Supermicro chassis. Big pools, hot-swap, rack.
How we picked, and what the lab showed
Every spec here was checked against the manufacturer datasheet and a live retailer listing, not pulled from memory, because hardware drifts and a NAS build punishes a wrong assumption harder than most. The choices that actually shape these lists came out of testing ZFS on our own hardware, not from a spec sheet.
The clearest result was on synchronous writes. On our test pool, ZFS handled roughly 334,000 IOPS with sync=standard but collapsed to about 2,041 IOPS with sync=always, a penalty of roughly 160x once every write had to hit stable storage before the pool acknowledged it. The same pattern showed up in fsync throughput: an NVMe device cleared about 2,024 fsync operations per second where a SATA SSD managed 642. That gap is the entire argument for a separate log device, and it is also the reason a log device does nothing for the workload most home builders actually have. If you are writing bulk files over SMB, those writes are asynchronous and ZFS batches them in RAM regardless. A log device only earns its slot when the writes are synchronous, which means NFS datastores for virtual machines, iSCSI, or a database living on the pool. We come back to that on the SLOG section below.
The other non-negotiable is the drives themselves. ZFS rebuilds a degraded pool by resilvering, a long sequential-and-random write storm across the surviving disks, and SMR drives stall catastrophically under exactly that load. We have watched an SMR resilver crawl for days where a CMR pool finished in hours. So only CMR and enterprise drives go in a ZFS pool, every time, and ECC memory backs the whole thing because ZFS trusts RAM to be correct. The platform decides which kind of ECC you buy, and that single distinction separates the consumer tiers from the server tier more than anything else on the parts list.
Build comparison: which ZFS server fits you
| Spec | 4-bay starter | 8-bay tower | 24-bay rackmount |
|---|---|---|---|
| Platform | AM5, Ryzen 5 7600 | AM5, Ryzen 7 7700 | EPYC SP3 7302 (used) |
| ECC type | UDIMM (unbuffered) | UDIMM (unbuffered) | RDIMM (registered) |
| RAM | 32GB | 64GB | 128GB |
| Data bays | 4 (board SATA) | 8 (IT-mode HBA) | 24 (SAS expander) |
| Network | 2x 1GbE (10GbE optional) | 2x 10GbE onboard | 2x 10GbE onboard |
| Form factor | Compact cube | Compact tower | 4U rackmount |
| Price (excl. drives) | ~$2,060 to $2,335 | ~$3,260 to $3,485 | n/a (mostly used market) |
| Best for | Backups, media, learning | Apps, VMs, the all-rounder | Big pools, hot-swap, rack |
Quiet 4-bay starter: a low-power ECC NAS for backups and media
The case sets the footprint of this build, so it comes first.
This is a quiet box that draws very little power and sits in a cupboard holding your backups and a media library. It runs on a Ryzen 5 7600, a 65-watt chip with integrated graphics so the build needs no GPU, and an ASRock Rack board that puts ECC UDIMM and IPMI on a small consumer-priced platform. The boot mirror goes on the board’s two M.2 slots, which keeps all four SATA ports free for spinning disks. Build this if you want to learn ZFS on real hardware without a server budget. The case is the one platform change since September: the Fractal Node 804 this build used went out of stock on Amazon, so it now sits in a Jonsbo N6, which takes the same micro-ATX board and ATX power supply and adds hot-swap trays, two 120mm drive fans and a three-speed fan switch.
Skip it if you already know you want more than four or five drives, or you need 10GbE on day one. The plain board ships with dual 1GbE, and going past four data drives means adding a SATA HBA. Both point you at the tower instead.
| Component | Pick | Approx price (Oct 2026) |
|---|---|---|
| CPU | AMD Ryzen 5 7600 (65W, iGPU) lead time | $220-$245 |
| Motherboard | ASRock Rack B650D4U (micro-ATX, ECC UDIMM, IPMI) Newegg/Best Buy | $300-$420 retailer, not re-priced |
| RAM | NEMIX 16GB DDR5 ECC UDIMM (buy two = 32GB) | $1,120-$1,200 |
| Boot mirror | WD Green SN350 250GB NVMe (buy two) 40 TBW | $155-$175 |
| Case | Jonsbo N6 (micro-ATX, 9 hot-swap bays) | $170-$180 |
| PSU | Seasonic CORE GX-650 (650W, ATX 3.1) | $95-$115 |
| Example HDD (separate) | WD Red Plus 6TB (CMR) | $290-$310/drive |
Three things on this build are worth a sentence. The ASRock Rack board is the part that makes ECC work here, because the 7600 supports ECC only when the board enables it, and this one does: ASRock Rack’s spec sheet lists DDR5 288-pin ECC and non-ECC UDIMM support explicitly. The modules linked above are NEMIX, a third-party module house rather than a name on ASRock Rack’s qualified vendor list (the tower’s A-Tech sticks are the same kind of part), which is the trade you make for them being in stock at all right now. The power supply is the same Seasonic CORE GX-650 the tower uses. 650W is still far more than an idle draw in the tens of watts needs, but it is an ATX 3.1, 80 Plus Gold, fully modular unit with a seven-year warranty and a normal in-stock listing, and the 750W Corsair RM750e this table named before is down to scarce and renewed listings. And the plain B650D4U has dual 1GbE, not 10GbE. If you want 10GbE in this footprint, the B650D4U-2L2T/BCM variant adds onboard dual 10GbE for about $515 from US resellers such as SabrePC (it has dropped off Amazon), which pushes the build past the top of its budget but saves a PCIe slot. Check the M.2 count before you swap, though: the 10GbE variant has one M.2 socket where the plain board has two, so the two-drive NVMe boot mirror in the table becomes one NVMe plus one SATA, and that costs you one of the four SATA ports this build keeps free for disks. A four-drive RAIDZ1 of 6TB CMR disks adds about $1,200 on top of the figures above, at what those drives cost now.
Eight-bay tower: 64GB ECC, an IT-mode HBA, and 10GbE
Eight hot-swap bays in one quiet tower are what separate this build from the starter.
This is the build most homelabbers should buy. It is an eight-bay box with 64GB of ECC, an IT-mode HBA feeding raw disks to ZFS, and onboard dual 10GbE so the network keeps up with the pool. The board is the centerpiece: an ASRock Rack server board that lists ECC UDIMM support on its own spec sheet, carries IPMI for headless management, and folds the 10GbE in so there is no separate NIC to buy or fit. Run apps, containers, and a few VMs straight off the pool, not just file shares.
Skip it if four bays is genuinely all you will ever need, in which case the starter saves you money, or if you are heading for 100TB and a rack, where the 24-bay build is the honest answer. For the broad middle, this is the one.
| Component | Pick | Approx price (Oct 2026) |
|---|---|---|
| CPU | AMD Ryzen 7 7700 (8C/16T, 65W) | $320-$350 |
| Motherboard | ASRock Rack B650D4U-2L2T/BCM (ECC UDIMM, IPMI, 2x 10GbE) SabrePC, not on Amazon | about $515 |
| RAM | A-Tech 32GB DDR5 ECC UDIMM (buy two = 64GB) | $1,690-$1,780 |
| HBA (IT mode) | LSI 9300-8i (SAS3008, pre-flashed IT) | $95-$115 |
| Network | Onboard dual 10GbE (none to buy) | $0 |
| Boot mirror | Samsung PM883 240GB (buy two, enterprise SATA with PLP) EOL | $320-$350 |
| SLOG (optional, PLP) | Intel Optane P1600X 118GB used market | used market, no new stock; not in the total |
| Case | SilverStone CS382 (8-bay hot-swap, micro-ATX) | $225-$260 |
| PSU | Seasonic CORE GX-650 (650W, ATX 3.1) | $95-$115 |
| Example HDD (separate) | WD Red Pro 16TB (CMR) | $680-$720/drive |
Where this build’s cost falls or rises is the memory. ECC UDIMM runs about $850 for a 32GB module now, so the 64GB in the table, as two 32GB sticks, is around $1,700, which is more than the board, CPU, case and power supply cost together. Two larger modules are also the better layout here: the board drops to 3600 MT/s with all four slots populated, so filling two slots keeps the faster memory; growing to four sticks later brings the 3600 MT/s drop back. If that still stings, start with one 32GB module and add the second later; ZFS is happy with 32GB on a pool this size, though a single stick runs on one memory channel until its partner goes in. The HBA needs a word too: buy a card whose provenance you can check, because counterfeit SAS3008 chips on the cheapest listings are a known source of dropped drives. The card linked above is a marketplace seller’s pre-flashed LSI, which is the normal way these are sold now; if you would rather not trust a reflash, buy an unflashed Broadcom or Supermicro card and flash it yourself. The chassis change is deliberate. This build used a Fractal Define 7, which ships with six drive trays, and the Type-B trays needed to reach eight are a separate part that is out of stock everywhere right now, so the eight-bay build could not actually be built. The SilverStone CS382 gives you eight hot-swap bays with nothing to source separately, takes the same micro-ATX board, and lands in the same place on price. Its backplane presents eight discrete SAS or SATA ports rather than a single mini-SAS socket, so cable it with two SFF-8643 to four-SATA forward-breakout cables from the 9300-8i’s two internal ports. Two smaller details from SilverStone’s spec sheet are worth knowing before you order. The backplane is powered through two 4-pin peripheral and two 15-pin SATA inlets, and SilverStone does not document the two as interchangeable, so have Molex leads available rather than assume a SATA-only supply will do. And the case takes two internal 2.5-inch drives, which is where the boot mirror goes rather than into two of the eight hot-swap bays. If you already own a Define 7, it still works as a six-drive box today, and as an eight-drive box whenever Fractal’s trays come back.
Rackmount beast: used EPYC and a 24-bay hot-swap chassis
The chassis defines this build: 24 hot-swap bays behind a single backplane.
This is the build for someone who already knows they want a big pool and a rack, and who is comfortable buying used server gear. The platform is a second-hand AMD EPYC 7302 on an ASRock Rack ROMED8-2T, which brings 128 PCIe lanes, eight-channel registered ECC, and onboard dual 10GbE. The centerpiece is a used Supermicro CSE-846, a 4U chassis with 24 hot-swap bays and a single SAS3 expander backplane, so one cable from one HBA drives all 24 disks with staggered spin-up. Where this falls short is noise and power: it is a rackmount box with screaming fans and a three-figure idle draw, not something for a living room.
Skip it if you are not ready to source parts from eBay and match an HBA to a backplane, or if quiet matters. The tower does almost everything this does, quietly, up to eight drives.
| Component | Pick | Approx price (Oct 2026) |
|---|---|---|
| CPU | AMD EPYC 7302 (16C/32T, SP3) used/eBay | used market, not re-priced |
| Motherboard | ASRock Rack ROMED8-2T/BCM (onboard dual 10GbE) used/eBay | used market, no new stock |
| ECC RAM | 4x Samsung 32GB DDR4-2933 RDIMM (128GB) used/eBay | used market, not re-priced |
| HBA (IT mode) | LSI 9305-16i, or 9305-8i with the expander chassis used/eBay | used market, not re-priced |
| Chassis + PSU | Supermicro CSE-846 24-bay + SAS3 expander + dual PSU used/eBay | used market, not re-priced |
| Network | Onboard dual 10GbE (Broadcom or Intel, by variant) | $0 |
| Boot mirror | Samsung PM883 240GB (buy two, enterprise SATA with PLP) EOL | $320-$350 |
| SLOG (optional) | Intel Optane P1600X 118GB used market | used market, no new stock |
| Example HDD (separate) | Seagate IronWolf Pro 20TB (CMR) low stock | about $860-$900/drive |
The one detail that trips people on this tier is matching the HBA to the chassis. With the CSE-846’s single-expander backplane, one SFF-8643 link feeds all 24 bays, so an eight-port 9305-8i is enough and cheaper than the 16-port card. You only need the 16-port 9305-16i for a direct-attach chassis with no backplane, where every port has to be cabled. Two more notes on the platform: the ECC here is registered RDIMM, not the unbuffered UDIMM of the smaller builds, because EPYC SP3 requires it, so do not buy the wrong kind. And watch for vendor-locked EPYC chips on used listings; buy one explicitly marked unlocked.
Shop the capacity rather than the model on the drives. The same 20TB class runs from about $830 for an IronWolf Pro to over $900 for an Exos X20 across live listings, so which listing you land on moves the cost of a populated pool more than most of the layout decisions you are weighing.
The two boot SSDs need somewhere to live that is not a hot-swap bay: most CSE-846 revisions have no internal 2.5-inch mounts, so budget for a rear or floor bracket kit, or accept that two of the 24 bays go to boot and the pool is a 22-drive pool. The chassis usually ships with dual redundant 80+ Platinum supplies, and for a fully loaded 24-bay pool the 1280-watt pair is the safer choice over the 920-watt. One board note: the base ROMED8-2T carries Intel X550 10GbE and the /BCM variant a Broadcom BCM57416. Both are supported by TrueNAS SCALE, so buy whichever the used listing offers and confirm which NIC it ships. Neither is sold new on Amazon any more: the only listings there are third-party dropshippers asking $1,580 and up for a board whose real used price is well under $700, so eBay and the server-surplus sellers are where this one comes from.
Pick CMR drives, never SMR
This is the rule that matters more than any board or CPU choice. ZFS resilvers a degraded pool with a sustained write storm, and shingled (SMR) drives stall under that load badly enough to drag a resilver from hours into days, sometimes long enough that a second drive fails first. Only CMR and enterprise drives belong in a ZFS pool. WD Red Plus and Red Pro are CMR; the bare WD Red at small capacities is SMR and is the classic trap. Seagate IronWolf Pro and Exos are CMR. We cover the specifics, including how to confirm a model’s recording technology before you buy, in the CMR versus SMR explainer and the wider NAS hard drive guide. Match the drives to the build above and you will not be the person posting a frozen resilver to a forum at 3am.
ECC RAM: UDIMM on consumer, RDIMM on server
ZFS assumes the memory holding your data is correct, which is why ECC is strongly recommended for any pool you care about. The catch that bites builders is that the two ECC form factors are not interchangeable, and the platform dictates which one you buy. Consumer AM5, the starter and tower builds here, takes unbuffered ECC, called UDIMM. Server EPYC and Xeon, the rackmount build, takes registered ECC, called RDIMM. Putting RDIMM in an AM5 board, or UDIMM in an EPYC board, simply will not post. So the rule is the platform decides: a Ryzen build means ECC UDIMM, an EPYC build means ECC RDIMM, and you never cross the two. As for capacity, 1GB of RAM per usable TB is the old rule of thumb, but for a home pool that is conservative; 32GB comfortably runs a small pool, 64GB suits an apps-and-VMs box, and you only need 128GB once the pool is large or you lean hard on deduplication.
Add a SLOG only if your writes are synchronous
A separate log device, the SLOG, is the most misunderstood part of a ZFS build. It does not cache reads, it does not speed up bulk file copies, and on the wrong workload it sits idle. What it does is absorb the ZFS intent log for synchronous writes, the ones where the client waits for confirmation that data is on stable storage before it continues. Our lab numbers put hard figures on that: the roughly 160x throughput gap between sync=standard and sync=always is exactly the cost a PLP-backed SLOG buys back, and only for sync writes.
That means a SLOG is worth fitting if you serve NFS datastores to a hypervisor, run iSCSI, or host a database on the pool. If your NAS mostly stores files over SMB, those writes are asynchronous and a SLOG changes nothing, so skip it and spend the money on drives or RAM. When you do add one, it must have power-loss protection, because a log device that loses in-flight writes on a power cut defeats its own purpose. Intel Optane was the favorite here for its low latency, and where Optane is now hard to find, a used enterprise NVMe or SATA SSD with explicit PLP is the substitute. The same power-loss-protection thinking is why the boot and special devices on the bigger builds use enterprise SSDs, and it pairs with a good UPS, which we cover in the homelab UPS guide.
Lay out the ZFS pool before you install
Decide the vdev layout before you push a single drive into a bay, because in ZFS you cannot remove a vdev from a pool once it is there, and the layout sets your fault tolerance for the life of the build. For a four-drive starter, a single RAIDZ1 vdev survives one drive failure and gives you three drives of usable space. For eight drives, RAIDZ2 is the standard call, surviving any two failures, which matters because the second failure often happens during the resilver of the first. Large pools on the 24-bay build are usually built from several RAIDZ2 vdevs of six to eight drives each rather than one enormous vdev, which keeps resilver times sane and IOPS reasonable.
Keep the boot devices off the data pool. Every build here mirrors the operating system onto two small SSDs that are separate from the spinning disks, so the pool is purely data and the OS can be rebuilt without touching it. On the rackmount and tower builds those boot SSDs are enterprise SATA with PLP; only the starter puts them on NVMe in the board’s two M.2 slots, which is why that build keeps all four SATA ports free for HDDs. If you are weighing NVMe against SATA for the supporting devices, the NVMe versus SATA comparison has the latency numbers.
Install TrueNAS SCALE and create the pool
TrueNAS SCALE installs from a USB stick onto the boot mirror, detects both SSDs, and lets you mirror the install across them during setup. Once it boots, build the pool in the web UI under Storage. That is the supported path on SCALE, and it matters more than it sounds: the middleware owns pool state, so a pool created behind its back does not appear as a managed pool and the scrub and SMART scheduling below will not be there to configure. If you want to see what the UI is doing, or you are building on plain OpenZFS rather than TrueNAS, this is the equivalent:
sudo zpool create -o ashift=12 -m /mnt/tank tank raidz2 \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0AAAAA \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0BBBBB \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0CCCCC \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0DDDDD \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0EEEEE \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0FFFFF \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0GGGGG \
/dev/disk/by-id/ata-WDC_WD80EFPX-68C4ZN0_WD-CA0HHHHHTwo details there earn their keep. The ashift=12 forces 4K sectors, which is correct for every modern drive and cannot be changed later. The -m /mnt/tank puts the mountpoint where TrueNAS expects it, because share paths are built under /mnt. Use the /dev/disk/by-id names rather than sdb through sdi: kernel letters are assigned in discovery order and move between reboots, and a pool that names them will eventually import against the wrong disks. Substitute your own serials: ls -l /dev/disk/by-id/ | grep -v part prints them. Confirm the layout and health before you put data on it:
zpool status tankThe pool and its raidz2-0 vdev should both read ONLINE with zeros in the READ, WRITE and CKSUM columns. Anything else, stop and work out why before you put data on it.
From there the UI handles datasets, SMB and NFS shares, snapshots, and scrub schedules. Set a monthly scrub and enable SMART tests so ZFS surfaces a failing drive before it takes the pool with it. If you are still deciding between rolling your own and a turnkey approach, the TrueNAS versus Unraid comparison weighs the two operating systems, and the home NAS roundup covers the prebuilt boxes if a kit is not your thing.
The mistakes that sink first-time ZFS builds
A handful of errors account for most of the unhappy NAS threads, and every one of them is avoidable at the parts-list stage. Putting SMR drives in a pool is the worst, because it only shows up when a drive fails and the resilver will not finish. Buying a RAID card instead of an IT-mode HBA is next; ZFS wants raw disks, and a RAID controller hiding the drives behind its own logic blinds ZFS to the very errors it exists to catch. Mixing ECC types, RDIMM in a consumer board or UDIMM in a server board, wastes money on parts that will not post. Undersizing the power supply so eight or twenty drives cannot spin up together is a quieter failure that shows up as random boot problems. And adding a SLOG to an SMB-only box spends money for zero benefit. Match the parts to one of the three builds above and you sidestep all of them.
The network is the last thing people underspend. A pool that can sustain several hundred MB/s is wasted behind a single 1GbE link, which caps you at about 110 MB/s. The tower and rackmount builds put 10GbE on the board for exactly this reason, and the 10GbE switch guide covers the other half of that link.
Which build should you actually buy?
Strip away the parts lists and the choice comes down to three honest questions.
- Do you need four bays or fewer, want it quiet, and are learning ZFS? Build the Jonsbo N6 starter. It is the cheapest way onto real ECC hardware, and it will happily back up your machines and serve a media library for years.
- Do you want one box that stores files and runs apps and VMs, with room to grow to eight drives? Build the eight-bay CS382 tower. It is the build most people should buy, and the 10GbE and IT-mode HBA mean you will not rebuild it in a year.
- Do you already know you want 100TB-plus, have a rack, and are comfortable on the used market? Build the 24-bay rackmount. Nothing else gets you that many hot-swap bays per dollar, as long as you can live with the noise.
When you are unsure between two tiers, size for the drives you will own in three years, not today. Empty bays are cheap to leave open and expensive to add later, and a ZFS pool you sized correctly the first time is one you never have to tear down and rebuild. If you would rather start from the storage media, the server SSD guide and the NAS drive guide pair with any of these three builds.
