Best Mini PCs for Homelabs

A homelab mini PC is optimised for a completely different set of constraints than an AI machine, and confusing the two is the most common and most expensive mistake in this category.

An inference box wants memory bandwidth above everything else. A homelab node wants I/O, cores, low idle power and firmware that will let you pass a device through to a virtual machine. Those requirements barely overlap, and the machines that serve them well look nothing alike.

What a homelab node actually needs

Networking, and more of it than you think. The moment you run a NAS and a hypervisor on separate boxes, gigabit becomes the bottleneck for everything — backups, VM migration, dataset transfers. Two interfaces also let you separate management traffic from workload traffic, which stops a backup job from making your dashboards unreachable.

Low idle power. A homelab node runs continuously. This turns out to dominate the running cost of a lab, and it is worth doing the arithmetic before buying — see below.

Cores over clocks. Containers and virtual machines want thread count. Sustained all-core performance matters far more than peak single-core boost, and a chip that boosts hard for thirty seconds then settles is a poor fit for a machine that is always doing something.

Memory capacity, and ideally sockets. Virtualisation is memory-hungry. A machine with SO-DIMM slots can grow; one with soldered memory cannot.

Storage flexibility. Multiple NVMe slots let you separate boot, fast VM storage and bulk capacity without external enclosures.

Firmware that exposes IOMMU. The specification nobody checks until it fails — covered properly further down.

Notice that memory bandwidth, the specification that dominates AI machine selection, does not appear on that list at all.

Idle power is the whole budget

This is the calculation that should come first, and almost never does.

A lab node runs 8,760 hours a year. Every watt of idle draw costs 8.76 kWh annually, so the arithmetic is simple:

The gap between a mini PC and a used enterprise server is the last two rows, and over five years it is frequently larger than the price difference that made the server look attractive.

It also compounds. A three-node cluster at 40 W idle plus a NAS and a switch is a permanent 150 W load — around $250 or £390 a year before anything actually does work.

Two practical notes. Manufacturer idle figures are measured in ideal conditions with nothing installed; expect real figures a few watts higher. And a cheap energy monitor at the wall is worth buying before you scale a lab, because the difference between what you assume and what you measure is routinely a factor of two.

The N100 and N150 class: when it is exactly right

Intel’s N-series has taken over the entry tier of homelab hardware, and the advice about it tends to be either dismissive or over-enthusiastic. Both are wrong, because it is a genuinely excellent chip for a narrow set of jobs.

What it is: four efficiency cores, no hyper-threading, a 6 W base power figure, and — the specification that matters most — single-channel memory only. Officially 16 GB maximum, though many boards run more.

What it handles comfortably:

• A firewall or router — this is the standout use, particularly on a box with two 2.5 GbE ports
• DNS filtering, a reverse proxy, a VPN endpoint
• Home automation
• A handful of light containers
• A small NAS serving a couple of clients

Where it runs out, quickly:

• Multiple virtual machines. Four cores with no SMT means four threads total. Two VMs and a container stack will contend.
• Anything memory-bandwidth-bound. Single-channel DDR5 gives roughly 38 GB/s, which a dual-channel machine doubles.
• Transcoding at scale. The Quick Sync engine is capable and the rest of the chip is not, so a couple of simultaneous streams is the realistic ceiling.
• Language models. Not at any size worth using.

The honest framing: an N-series box is an appliance, not a hypervisor. Buy one to do one job continuously at six watts, and it is superb value. Buy one as your only lab node and you will replace it within a year.

The specifications

The recommendations

Best overall homelab node

The MS-01 is the machine that most homelab requirements converge on, and the networking is why.

Two 10 GbE SFP+ ports in a one-litre chassis is close to unique at this size and price. Add two 2.5 GbE copper ports and you have enough interfaces to separate management, storage and VM traffic properly — the sort of thing that stops being theoretical the first time a backup saturates your only link.

Then a PCIe 4.0 x16 slot (x8 electrical) that accepts a half-height card, three NVMe slots including one U.2-capable, and up to 96 GB of DDR5 across two SO-DIMM sockets. That is genuinely unusual expandability for the footprint, and the SO-DIMM sockets mean a memory decision you can revisit.

Fourteen cores handle a substantial container workload. Idle draw is around 15 W.

Two honest caveats: the fans are audible under sustained load, so it is not a living-room machine; and the 13th-generation mobile silicon is no longer current, though for virtualisation the I/O matters far more than the core architecture.

Best entry-level appliance

If what you actually need is a router, a firewall or a single always-on service, this is the right amount of machine and the rest of this page is over-specification.

Two 2.5 GbE interfaces at six to eight watts is exactly the shape of a network appliance. OPNsense or pfSense will route well beyond a gigabit on it, and it will do so silently for roughly the cost of a coffee a month.

Do not buy it as a virtualisation host. Four efficiency cores and single-channel memory are the constraints described above, and they arrive sooner than people expect.

Best low-power capable node

If your lab lives in an office rather than a cupboard, noise and idle power move up the list — and this is the machine that gives you real capability without either.

Twelve Zen 5 cores at around 10 W idle is an excellent basis for a containers-and-services box, and it is genuinely quiet. Wi-Fi 7 is useful for a node that is not near a switch.

The limitations are networking and expandability: 2.5 GbE only, and soldered memory. If your lab will grow toward 10-gigabit, this is a machine you will supplement rather than upgrade.

If the node will also run models

Worth considering when you want one machine to do both.

128 GB of unified memory runs 70B models and simultaneously gives you a very large pool for virtual machines. Sixteen Zen 5 cores handle containers comfortably. 5 GbE sits usefully between 2.5 and 10 gigabit.

The trade-offs against a dedicated node are real: no PCIe expansion slot, one fewer NVMe slot than the MS-01, and soldered memory. You are paying a premium for inference capability that a pure services node does not need — see Best Mini PCs for Local LLMs if that is actually the priority.

Virtualisation: the specifications that bite later

IOMMU, and why firmware matters more than silicon

If you intend to pass a physical device through to a virtual machine — a network card to a firewall VM, a GPU to a media server, an HBA to a storage VM — you need IOMMU (Intel calls it VT-d, AMD calls it AMD-Vi) working properly.

Every modern processor supports it. Not every mini PC firmware does, and this is where the category disappoints:

• Some boards do not expose the toggle in the BIOS at all.
• Many have poor IOMMU grouping, where a whole set of devices lands in one group and must be passed through together — which usually means you cannot pass any of them through, because one of them is your boot controller.

The workaround is an ACS override patch, which weakens the isolation IOMMU exists to provide. It works, and it is not something to build on if you have a choice.

Check before you buy. Search for the specific model plus “IOMMU groups” or “passthrough” and read what people actually got. This is the single most common source of homelab regret, and it is entirely invisible on a specification sheet.

ECC, honestly

Consumer mini PCs do not have ECC memory. Intel’s N-series does not support it, and AMD mobile parts that technically can almost never have it enabled on the board.

What ECC protects against is a single-bit memory error silently corrupting data in flight. For a machine running containers that can be rebuilt from a compose file, that risk is academic. For a storage host holding the only copy of something you care about, it is not — a bit flipped in memory gets written to disk, and checksums downstream will faithfully protect the corrupted version.

The honest position: the absence of ECC is not a reason to avoid mini PCs for a lab. It is a reason not to make a consumer mini PC the only place your irreplaceable data lives.

Nested virtualisation

If you plan to run a hypervisor inside a virtual machine — testing a Proxmox upgrade, running Docker Desktop inside a Windows VM, building a Kubernetes lab — you need nested virtualisation enabled. Modern Intel and AMD parts support it; it is off by default on most hypervisors and is a one-line change.

Storage inside the node

A virtualisation host writes far more than people expect, and for a reason that is not obvious.

Write amplification stacks. A 4 KB write inside a guest becomes a larger write on the guest filesystem, then a larger one again on the host filesystem, then a full NAND page programme on the SSD. Five to ten times amplification is normal. ZFS with its default record size makes it worse for small random writes.

Power-loss protection matters more than endurance. Enterprise SSDs carry capacitors that let in-flight writes complete during an outage. Consumer drives do not — which is why sync-heavy filesystems either run slowly on them or, configured carelessly, risk losing acknowledged writes. If you are running ZFS with synchronous writes, this is the specification to look at, not TBW.

A sensible layout in a three-slot machine:

More on drive selection in Best NVMe SSDs for AI Workloads, which covers sustained write behaviour and the CMR/SMR trap in detail.

Networking around the node

Why multiple interfaces. Once you have VLANs, one physical port carrying a trunk is workable but fragile — a misconfiguration takes management down with everything else. A dedicated management interface means you can always reach the box.

SFP+ or copper? A detail with real consequences at home:

That difference is why fanless 10-gigabit switches are almost always SFP+, and why a silent lab tends to end up on DAC cables.

And storage that can actually saturate the link:

Clustering, and why people buy three nodes

If you read homelab forums you will notice everyone owns three of something. There is a reason, and it may not apply to you.

Quorum. A Proxmox cluster makes decisions by majority vote. With two nodes there is no majority when they disagree, so a two-node cluster stops working the moment one node fails — the opposite of what a cluster is for. Three nodes is the minimum that survives losing one. (A two-node cluster plus a QDevice — a third vote from something as small as a Raspberry Pi — is a legitimate and much cheaper alternative.)

Ceph. Proxmox’s distributed storage wants three nodes minimum and, realistically, a dedicated 10-gigabit network. On gigabit it is slow enough to be miserable.

Whether you need any of this. Most home labs do not. High availability protects against unplanned hardware failure; it does nothing for the far more common cause of downtime, which is you changing something. A single well-backed-up node plus a documented rebuild is a better use of money for most people than three nodes and a Ceph cluster.

Build the cluster when you have a workload that genuinely cannot be down, not because the forum has one.

Noise and placement

Worth planning before rather than after.

Mini PCs are quiet at idle and most are audible under sustained load — small fans have to spin fast to move air. The MS-01 in particular is noticeable when it is working.

If the machine is going in a cupboard, remember that a cupboard has no airflow. A node that idles at 15 W will warm a small enclosed space enough to raise its own intake temperature, at which point the fans run faster permanently. A vent, or a gap, or a low-speed 120 mm fan, solves it.

Mini PC or used enterprise server?

The genuine alternative, and worth comparing honestly rather than dismissing.

Buy the server if you have somewhere it can be loud — a garage, a basement, a rack in an outbuilding — and you want many cores, lots of ECC memory and proper out-of-band management for very little money.

Buy the mini PC if the machine has to live near people. Silence and six watts are worth a great deal, and a lab that annoys the household gets switched off.

What to avoid

Machines with a single network interface. You will want to separate traffic sooner than you expect, and it cannot be added later.

Soldered memory, if your requirements are unsettled. Virtualisation memory needs grow. SO-DIMM sockets let you follow them.

Boards with no IOMMU option in firmware. Check before buying, not after.

N-series boxes as hypervisors. Excellent appliances, poor virtualisation hosts. The distinction is the whole of this category’s buying advice.

Buying AI capability a services node will never use. A Strix Halo machine costs several times a services node. If the lab is going to run Home Assistant and a media server, that money buys nothing at all.

Continue your research

• Best Mini PCs for Local LLMs — if the node also needs to run models
• Best NVMe SSDs for AI Workloads — drive selection, sustained writes and the CMR/SMR trap
• Best AI Workstations — the machine on the other end of the network
• Best Monitors for Developers — including the KVM that switches between both