Server Network Adapters (NICs) Explained: 1/10/25/100GbE, SFP28 vs RJ45 & OCP vs PCIe
How to choose a server network adapter — the speeds (1G to 100G), connectors (RJ45, SFP+, SFP28, QSFP28), form factors (PCIe, OCP 3.0, Dell rNDC, HPE FlexLOM), and offloads like RDMA/RoCE and SR-IOV that decide whether a NIC fits your server and your network. With real part numbers.
TL;DR — four things pick a server NIC
- Speed — 1G, 10G, 25G, 40G, 50G, or 100G. The data-center sweet spot has moved to 25G (servers) and 100G (uplinks/AI/storage).
- Connector — RJ45 (BASE-T copper), SFP+ (10G), SFP28 (25G), QSFP+ (40G), QSFP28 (100G). The NIC's cage must match your switch optics/DAC.
- Form factor — standard PCIe card, OCP 3.0 mezzanine, or vendor slots (Dell rNDC/bNDC, HPE FlexLOM/FLR). It has to physically fit the server.
- Offloads — RDMA/RoCE, SR-IOV, TCP/checksum offload. Critical for storage (NVMe-oF, vSAN), HPC, and AI; optional for general traffic.
Get speed + connector + form factor right and the card works; pick the offloads for your workload. (NVIDIA ConnectX adapters span 10-400GbE.)
---
Speed and connector go together
| Speed | Connector | Typical use |
|---|---|---|
| 1GbE | RJ45 | Management, legacy |
| 10GbE | RJ45 (BASE-T) or SFP+ | Mainstream server / uplink |
| 25GbE | SFP28 | Modern server access (the new 10G) |
| 40GbE | QSFP+ | Older uplinks/aggregation |
| 50/100GbE | QSFP28 | Uplinks, storage, AI/HPC |
A NIC's port cage is fixed: an SFP28 NIC takes 25G (and 10G) SFP modules/DAC; a QSFP28 NIC takes 100G (or breaks out to 4x25G). Match the NIC cage to your switch — see SFP vs SFP+ vs QSFP+ vs QSFP28 and 10G vs 25G vs 40G vs 100G server networking.
---
Form factor — the fit trap
A NIC has to seat in the slot your server actually has:
- PCIe (standard) — universal add-in card; the flexible default.
- OCP 3.0 — small mezzanine slot in modern servers; frees a PCIe slot.
- Dell rNDC / bNDC, HPE FlexLOM / FLR / OCP — vendor-specific mezzanine connectors.
A PCIe NIC won't fit an OCP slot and vice versa. Confirm the slot type for your exact server (this is the same lesson as the HP ProLiant network adapter guide).
---
Offloads: RDMA/RoCE, SR-IOV, and why they matter
- RDMA over Converged Ethernet (RoCE) lets the NIC move data directly to memory, bypassing the CPU — huge for NVMe-oF, vSAN, GPUDirect, and AI. RoCE needs a lossless network (PFC/ECN configured on the switches).
- SR-IOV presents virtual NIC functions to VMs/containers for near-bare-metal networking.
- TCP/checksum/LSO offload is standard and reduces CPU overhead on general traffic.
If your NIC is feeding GPUs or an all-flash storage fabric, prioritize an RDMA-capable adapter; for plain VM traffic, any modern 10/25G NIC is fine.
---
What you can buy (real inventory examples)
- 100GbE: Mellanox/NVIDIA ConnectX-5 dual-port 100Gb QSFP28 (Dell 9FTMY) — RDMA/RoCE for storage/AI.
- 25GbE: Broadcom BCM57414 dual-port 25Gb SFP28 (HPE P53862-B21).
- 10GbE: Broadcom 57810S dual-port 10GBASE-T (Dell HN10N), Intel X710-T4L quad 10GBASE-T (Dell FWH37).
- CNA / converged: QLogic QLE8262 dual 10Gb (Dell JHD51).
Compare the silicon vendors in ConnectX vs Intel vs Broadcom NICs.
---
FAQ
RJ45 or SFP28? RJ45 (BASE-T) reuses copper cabling and is simple for 10G; SFP28 (25G) with DAC/optics is the modern data-center choice and runs cooler at distance.
Do I need RDMA? For NVMe-oF, vSAN, or AI/GPU fabrics, yes. For general server traffic, no.
Will a 25G NIC work on a 10G switch? SFP28 NICs are backward compatible with 10G SFP+ — they'll link at 10G.
---
Pro Disk Network is an independent reseller of genuine NICs from NVIDIA/Mellanox, Intel, Broadcom, and others (not affiliated with those vendors). Next: ConnectX vs Intel vs Broadcom comparison. Related: Broadcom 5719/5720/57810 NIC guide.
Source: NVIDIA ConnectX Ethernet Adapters (speed range, RoCE/RDMA). Pricing/availability reflect Pro Disk Network US inventory.