NVMe-TCP vs. Fibre Channel vs. iSCSI: A Practical Decision

Anyone planning a new mid-market storage cluster today faces a decision that was clear five years ago: iSCSI for entry-level, Fibre Channel for production. In 2026 the landscape has shifted. NVMe over TCP is stable in TrueNAS 25.10 and Proxmox VE 9, 32Gb Fibre Channel remains the reliable constant in many data centres, and iSCSI is enjoying a late renaissance thanks to more efficient offload cards. We show how we weigh the three protocols against each other in customer projects and which setup fits which requirement.

This article is aimed at IT decision-makers asking: what do we deploy now so it carries us for the next seven years? We deliberately avoid quoting our own benchmark numbers — every environment is too different to derive reliable statements for you. Instead, we structure the discussion by requirement profile.

The three candidates at a glance

iSCSI transports SCSI commands over TCP/IP. It runs on any standard network card, has been production-ready in TrueNAS and Proxmox for over a decade, and comes with a mature multipath stack. The downside: SCSI is a sequential command protocol from the 1980s that cannot keep up with modern NVMe drives once queue depth grows.

Fibre Channel (FC) is a dedicated storage network with its own fabric, switches and cabling (typically OM4 multimode for 32Gb connections up to 100 metres). The protocol is lossless, deterministic and deeply embedded in enterprise environments. 32GFC is the current standard, 64GFC is available, but rarely economical in the mid-market.

NVMe over TCP (NVMe/TCP) transports NVMe commands natively over TCP/IP — without the SCSI detour. This removes the architectural brakes of the iSCSI stack. Multiqueue, deep command queues and parallel I/O are part of the protocol itself. Runs on regular Ethernet cards from 25 GbE upwards.

Setup profile TrueNAS: NVMe-TCP with a Proxmox cluster

For a typical greenfield build — three Proxmox nodes, one TrueNAS storage system with an NVMe pool — NVMe-TCP is our default recommendation. The cabling stays Ethernet, the switches do not need special FC certification, and configuration effort is manageable.

# TrueNAS SCALE 25.10: enable NVMe-TCP target
# Storage -> Sharing -> NVMe-oF -> Configure Service
# Portal on a dedicated storage VLAN, port 4420 (default)

# Proxmox VE 9 client side: nvme-cli
apt install nvme-cli
modprobe nvme-tcp

# Discover subsystem
nvme discover -t tcp -a 10.20.30.10 -s 4420

# Connect namespace
nvme connect -t tcp -n nqn.2011-06.com.truenas:pool01:vmdata \
  -a 10.20.30.10 -s 4420 -l 1800

Important: a dedicated storage VLAN, MTU 9000 on all participating ports, and at least two physically separated paths. The nvme-tcp client in kernel 6.8+ includes native multipath support — no multipathd overhead as with iSCSI. In our projects we see noticeably lower CPU load on the hypervisor side than with comparable iSCSI setups, especially with parallel VM workloads.

Setup profile Fibre Channel: when the fabric already exists

If you already run a Brocade or Cisco MDS fabric in your data centre, you do not throw it away lightly. The typical situation: an existing SAN with HPE Alletra or NetApp AFF that still has three to five years of runtime left is extended with a Proxmox cluster. Then FC HBAs (QLogic QLE2772 or Emulex LPe35002 for 32GFC) are the right choice.

Proxmox supports FC LUNs via the standard multipath stack. The LUN appears as a block device, LVM-thin or a shared filesystem such as OCFS2 sits on top. The advantage: the latency characteristics of FC are predictable over years, the fabric does not buffer packets, and zoning provides hard isolation.

For pure greenfield projects in the mid-market, FC rarely pays off any more. Two FC switches, HBAs per host, transceivers and cabling quickly add up to a five-figure premium over a 25 GbE Ethernet solution. The benefit is real but expensive.

Setup profile iSCSI: the pragmatic path

iSCSI remains the right choice when extending an existing infrastructure without storage VLAN separation, when your storage targets do not yet speak NVMe-TCP, or when you need to talk to older hypervisors (VMware ESXi 7.x, older Hyper-V clusters) that do not fully support NVMe-TCP.

TrueNAS has always offered iSCSI as a sharing option. On the Proxmox side, setup via the GUI takes a few minutes. Important: consistent use of CHAP authentication, dedicated portals per VLAN, and iscsid.conf tuning under heavy parallel workloads (node.session.cmds_max, node.session.queue_depth).

The honest view: iSCSI is functional but technically at the end of its development track. For new clusters it is the fallback option, not the default.

Decision matrix

Criterion	iSCSI	Fibre Channel	NVMe-TCP
New network investment	none	high (fabric, HBAs)	medium (25/100 GbE)
Reuse existing FC fabric	no	yes	no
Latency determinism	medium	very high	high
Multipath maturity	mature	mature	native to protocol
Host CPU overhead	high	low	low
VM density scalability	limited	high	very high
Compatibility with older hypervisors	very good	very good	limited
Greenfield recommendation 2026	fallback	rarely economical	default

The matrix does not replace a conversation about your specific environment, but it shows the direction. Anyone planning today for seven years should have good reasons not to choose NVMe-TCP.

Cabling, switching and the truth about 25 GbE

For NVMe-TCP in the mid-market, 25 GbE has established itself as the sweet spot. Switches from Mikrotik, Ubiquiti Enterprise or the MikroTik CRS520 deliver the required port density at prices that were unthinkable three years ago. What matters is SR-IOV-capable NIC hardware (Mellanox/NVIDIA ConnectX-6 or -7, Intel E810) and consistent separation of storage and management traffic at VLAN level — better still, physically.

The jump to 100 GbE is worthwhile if you push more than four nodes against the storage target simultaneously. Below that it is overkill — a single NVMe session rarely saturates 25 GbE. If you are unsure which setup fits your load profile, take a look at the current Proxmox and OPNsense configurations we use in customer projects.

Conclusion: the new default is NVMe-TCP

For greenfield projects in the mid-market, NVMe-TCP is the right choice in 2026. The protocol architecture fits modern NVMe pools, network costs stay within the Ethernet budget, and operational complexity is lower than with FC. Fibre Channel keeps its place with existing fabrics and requirements for absolute latency determinism. iSCSI is the compatibility bridge, no longer the standard.

DATAZONE supports you in the design, procurement and operation of TrueNAS and Proxmox storage clusters — from the first capacity plan to the multipath setup in production. Talk to us at /en/kontakt/ if you do not want to leave your next storage decision to chance.