basforge

Search titles, text and the “If it doesn’t work” cases.

    RU
    OverviewHomelabProxmox

    Hardware for Proxmox: CPU, memory, disks and network cards

    Proxmox VE installs on almost any 64-bit computer with VT-x or AMD-V, even a decade-old PC. The limits are elsewhere: memory shared by the guests and the ZFS cache, and SSDs — consumer drives without power-loss protection are slow at synchronous writes and wear out faster under ZFS. Intel I219 and some Realtek network cards need one line of configuration.

    What the CPU must support

    Proxmox VE 9 installs on 64-bit x86 — Intel 64 or AMD64. For virtual machines the CPU must support hardware virtualization: Intel VT-x or AMD-V. Almost every CPU since 2006 has it, but it’s often disabled in the BIOS out of the box. LXC containers don’t need it; VMs require it.

    What What for How to check
    VT-x or AMD-V (flags vmx, svm) any virtual machine grep -cE 'vmx|svm' /proc/cpuinfo — above zero
    VT-d or AMD-Vi (IOMMU) passing a GPU, disk controller or network card into a VM dmesg | grep -e DMAR -e IOMMU — lines about IOMMU being enabled
    AVX2 (x86-64-v3 level) guests built for v3: RHEL 10, Rocky Linux 10 grep -c avx2 /proc/cpuinfo
    AES-NI encryption in guests, the default x86-64-v2-AES CPU type grep -c aes /proc/cpuinfo

    You don’t need many cores: a guest’s virtual core is a thread on the host, and idle guests barely touch the CPU. The minimum in the Proxmox documentation is a 64-bit CPU with VT-x or AMD-V and 1 GB of memory plus memory for the guests — for evaluation. The 9 branch also runs on a dual-core CPU over ten years old with four threads, and that’s enough for a dozen light guests. Such a host hits the limits of memory and disks; the CPU matters when guests compute: transcode video, build code, compress backups.

    Since August 2026 there’s 9.2 for ARM (arm64), but officially only NVIDIA Grace and Vera servers; other hardware with ARMv9-A and UEFI is “best effort”. A Raspberry Pi won’t do: it describes its hardware with a device tree, while Proxmox on ARM needs UEFI and ACPI.

    How much memory

    Memory is divided strictly: what’s given to a running VM, the host won’t hand to others until the guest returns it. So count it in advance, guest by guest:

    Who How much
    Proxmox VE itself, with two small containers about 1.6 GB
    ZFS cache (ARC), if the host is on ZFS 10 % of memory, but no more than 16 GiB — that’s what the installer sets since 8.1
    A container with one service: Pi-hole, a small site give 128–512 MB; idle Pi-hole takes about 40 MB
    An Ubuntu server with Docker and a dozen containers 2–4 GB
    Home Assistant OS 2 GB — the minimum from the Home Assistant documentation
    Windows 11 from 4 GB, plus TPM 2.0 and UEFI — Proxmox provides virtual ones

    Hence a rule of thumb: 8 GB is enough for a couple of containers and one VM; 16 GB holds three or four VMs and several containers if the host is on ext4; ZFS and Windows guests ask for 32 GB. The Proxmox documentation has its own rule for ZFS: 8 GB to start with and about 1 GB of memory per TB of used space.

    Two things stretch memory, but don’t count on them:

    • Balloon. A driver in a Linux guest returns memory the guest isn’t using to the host. It works when the VM has both a minimum and a maximum memory set and the guest has the virtio_balloon driver — in Linux it’s in the kernel.
    • KSM. The host kernel looks for identical memory pages across VMs and stores them once. The ksmtuned service turns it on when more than 80 % of memory is in use. Ten identical Ubuntu VMs save hundreds of MB with it; different systems save almost nothing.

    ECC memory catches and corrects single-bit errors. It’s recommended for ZFS, but without it ZFS is just as safe as any other file system without ECC. At home it’s a question of what the board supports, not a requirement.

    Disks

    Disks are what most decides how Proxmox will run. The host itself writes constantly: graph statistics, /etc/pve in an SQLite database, logs. Guests add their own databases and logs. Hence the rules:

    • System on an SSD. On an HDD the host and guests work, but every action in the UI and every VM boot waits for the disk. A USB stick won’t do for the system: constant writes will kill it within months.
    • No hardware RAID under ZFS. ZFS must see the disks directly: the Proxmox documentation says outright that neither ZFS nor Ceph is compatible with a RAID controller. A controller in HBA mode (IT mode) is fine.
    • Guest SSDs with write endurance to spare. Endurance is listed in the spec as TBW — how many terabytes can be written under warranty. Consumer 500 GB models are 100–300 TB; 1 TB server drives with PLP start at 1500 TB.
    • PLP — power-loss protection. Capacitors on the SSD board flush the cache if power is lost. So such a drive acknowledges a synchronous write immediately, from the cache. Without PLP every synchronous write waits for the flash: databases and ZFS with synchronous writes run several times slower.
    • QLC — not for guests. QLC memory is fast while writes go into the SLC cache (a few tens of GB); beyond that the speed drops to tens of MB/s. Fine for ISOs, templates and backups.
    • HDD — for data and backups. A media library, files, a backup archive. VM disks on an HDD work, but are slow at random access.

    How Proxmox storages are laid out on these disks and when to use ZFS — Proxmox storage.

    How many disks

    How many How to lay them out
    One SSD the installer with ext4: the system and local-lvm for guests on one disk. Backups go to another machine or a NAS
    Two SSDs ZFS RAID1 (mirror) in the installer: one disk dies — the host keeps running, swap the disk live
    SSD and HDD system and guests on the SSD, the HDD as storage for backups and data
    Two SSDs and data disks a ZFS mirror for the system and guests, a separate ZFS pool of HDDs for data

    Network

    One gigabit card is enough: the host and all guests sit on one home network through the vmbr0 bridge (Proxmox networking). A second card is needed if a VM runs a router (one port to the provider, the other to the home) or the storage network is kept separate from the rest.

    Some cards have known quirks:

    • Intel I217 and I219 (driver e1000e, built into lots of boards and mini PCs) hang under load with the message Detected Hardware Unit Hang, and the network drops for a minute. The fix is turning off the TSO and GSO offloads (see “If it doesn’t work”).
    • Realtek RTL8125 (2.5 Gbit/s) works on the stock r8169 in recent kernels. Install the separate r8125 driver only if the stock one loses the link: it’s built through DKMS for every new kernel, and without the proxmox-headers package a kernel update leaves the host without network.
    • Intel I225 and I226 (2.5 Gbit/s) — early I225 revisions lost the link; if the card is on the board, check the revision.
    • USB adapters work; the kernel names them by MAC: enx74d4…. Use them for constant load only if there’s nothing else.

    A card’s name can change after a kernel update — and the vmbr0 bridge loses its port. Since 9.1 the installer offers to pin the names: the cards are then called nic0, nic1 and don’t change.

    What to get

    Most common at home

    Mini PC

    Intel N100 or N150, recent Core i3/i5, Ryzen 5. About 6–14 W idle, silent. Downsides: one SSD slot (sometimes two), often one network card, and the N100 takes a single memory module of up to 16–32 GB.

    An old PC or office desktop

    Office Dell OptiPlex, HP EliteDesk, Lenovo ThinkCentre of generations 6–10: VT-x and VT-d, several disks, a slot for a second network card. 15–30 W idle.

    A decommissioned server

    Lots of ECC memory, IPMI for a remote console, a drive cage. The price is noise and 60–150 W idle: year-round it shows on the electricity bill.

    An x86 NAS

    Several disks and an x86 CPU. Proxmox goes on instead of the NAS OS if the vendor allows it; the data disks go to a VM running the NAS or stay in a ZFS pool on the host.

    Before buying

    1. The CPU supports VT-x and VT-d (on AMD — AMD-V and AMD-Vi), and they can be enabled in the BIOS.
    2. It takes at least 16 GB of memory, better 32 GB or more.
    3. For the system and guests — a TLC SSD with TBW in the spec. For a ZFS mirror — two slots.
    4. The network card is a known Intel or Realtek model; two of them for a router in a VM.
    5. If you need GPU passthrough, the motherboard puts devices into separate IOMMU groups (check reviews of the specific board).

    If it doesn’t work

    A VM doesn’t start: KVM virtualisation configured, but not available

    Cause
    Hardware virtualization is disabled in the BIOS: Intel VT-x (Virtualization Technology) or AMD SVM Mode.
    Fix
    Enable it in the BIOS. Check on the host: grep -cE 'vmx|svm' /proc/cpuinfo — a number above zero.

    Rocky Linux 10 or RHEL 10 in a VM crashes right after the boot loader: Fatal glibc error: CPU does not support x86-64-v3

    Cause
    The VM’s default CPU type is x86-64-v2-AES, without AVX2. The 10 branch of these systems is built for x86-64-v3.
    Fix
    On the VM, Hardware › Processors › Type — x86-64-v3 or host; the host CPU must support AVX2 (Intel from Haswell, AMD from Excavator).

    The host network drops for a minute, and journalctl -k shows e1000e … Detected Hardware Unit Hang

    Cause
    The onboard Intel I217 and I219 cards on the e1000e driver hang under load with TSO and GSO offloads enabled.
    Fix
    Turn them off: ethtool -K eno1 tso off gso off, and to survive a reboot add post-up ethtool -K eno1 tso off gso off to the card in /etc/network/interfaces. Use your card’s name.

    VMs are slow on writes, iowait is high, though the SSD is new

    Cause
    A consumer QLC SSD filled its SLC cache and writes at tens of MB/s. ZFS and synchronous database writes bring that moment closer.
    Fix
    fio or copying a large file shows the speed dropping after a few GB. For guests, use TLC SSDs, preferably with power-loss protection (PLP); QLC is for ISOs and backups.

    The SSD’s Wearout in Disks grows by percents a month

    Cause
    ZFS, guest logs, databases and Proxmox itself write to the disk constantly. A cheap SSD rated for 100–300 TB written (TBW) wears out under such load in years, not in a decade.
    Fix
    See how much has been written: smartctl -a /dev/sdX (Total_LBAs_Written, for NVMe Data Units Written) against the TBW from the spec. Move heavy writes to a separate disk or an SSD with TBW to spare.