Building a Proxmox Host on Refurbished Hardware

Proxmox VE is one of the most rewarding things to run on refurbished enterprise hardware. The hypervisor itself is light, KVM does not care that your Xeon is a few generations old, and the platform features that matter most — ECC memory, lots of DIMM slots, out-of-band management, hot-swap bays — are exactly what enterprise servers have always shipped with. What decides whether the host feels fast is not the age of the silicon; it is how the box is configured.

CPU: cores are cheap, so buy the mid-range

Virtualization is the workload where extra cores pay off most directly, and refurbished pricing makes them nearly free. On 13th-generation Dell (R630, R730) and HPE Gen9 the sweet spot sits in the Xeon E5-2650 v4 to E5-2680 v4 band: 12 to 14 cores per socket, sensible clocks, and far cheaper per core than the top-bin 22-core parts. A dual-socket host in that range gives 24–28 physical cores — enough to run twenty ordinary VMs without scheduling pressure.

One practical note: Proxmox exposes NUMA topology to guests. Two sockets are great for capacity, but keep individual VMs inside one socket's core count where you can; a VM that spans NUMA nodes pays a latency tax on every remote memory access.

Memory: fill channels before chasing capacity

Both Haswell/Broadwell Xeons and their HPE equivalents run four memory channels per socket. Eight identical RDIMMs (four per socket) is the minimum layout that uses full bandwidth; 8× 32GB for 256GB is the classic Proxmox host configuration. ZFS wants memory too — budget roughly 1GB of RAM per 1TB of pool as a starting point, more if you enable deduplication (or better: do not enable deduplication).

Storage: the controller matters more than the disks

If you plan to use ZFS — and on Proxmox you probably should — the storage controller must present disks directly. That means an HBA in IT mode or a RAID controller switched to HBA/passthrough mode. Dell's H330 flashes cleanly to passthrough; the H730 offers a native HBA mode. On HPE, the B140i software mode should be avoided entirely and the P440ar set to HBA mode. A hardware-RAID virtual disk under ZFS hides SMART data and defeats ZFS's self-healing — the classic mistake in first builds.

For the pool itself: a mirrored pair of enterprise SATA or SAS SSDs for VM disks covers most labs and small businesses. Spinning drives still make sense for bulk secondary datasets, and an 8SFF chassis with two SSD mirrors plus four HDDs in RAIDZ is a tidy single-box layout.

Network: 10GbE is the upgrade you will feel

Live migration, backups to a NAS and VM restores all move whole disk images. Refurbished 10GbE cards — Intel X520/X710, Mellanox ConnectX-3/4 — cost little and are natively supported by the Proxmox kernel. If your switch has SFP+ ports, DAC cables keep it simple and cool.

A concrete example

A Dell R730 8SFF with 2× E5-2660 v4 (28 cores total), 256GB in eight 32GB RDIMMs, an H730 in HBA mode, two 960GB SSDs mirrored for VMs, four 4TB SAS drives in RAIDZ for bulk data and a dual-port X520 covers a serious home lab or a small company's entire virtual estate — usually for less than the price of one new entry-level tower.

The same recipe scales down to an R630 for a quieter 1U footprint, or up to an R730xd when the storage pool is the point. Whichever chassis you pick, the order of operations is the same: channels populated, controller in passthrough, ECC everywhere, and 10GbE if anything else on the network can speak it.