Hardware compatibility & sizing · NICs, drivers, XDP modes · interactive calculator

Infrastructure Brief · Hardware & Sizing

Which card, which driver, how many boxes.

BNGSOFT runs on commodity x86 with no appliance and no vendor lock-in — which shifts the hardware decision onto you. This brief lists the cards and drivers, states which XDP mode each one actually gets, and gives you a calculator that sizes against your traffic rather than a vendor’s assumption.

The number that moves sizing most is not ours — it is your busy-hour megabits per subscriber. So it is an input, not a constant.

Size it against your own traffic

Subscribers

Active subscribers at busy hour

Busy-hour Mbps / subscriber

Take this from your own netflow, not from a datasheet

Average frame size

Residential mixed traffic typically lands near IMIX

Redundancy

N+1 adds one spare box to the count

—

Throughput needed

—

Packet rate needed

—

Boxes

—

Suggested build

—

What this calculator does and does not do. It computes throughput and packet rate from your inputs, checks both against the 50 Mpps engine ceiling and against installed port capacity, and tells you which of the two binds. It deliberately does not estimate CPU cores. Across our own 68-gateway fleet, core consumption correlated with subscriber count at −0.16 and with packet rate at −0.13 — that is, not at all. Any core figure derived from that data would be invented precision. Core sizing comes out of a proof of concept on your traffic, not out of a web form.

Network cards

These are the cards behind the reference builds. Prices are indicative and move constantly; power figures are typical draw, not maximum.

Adapter	Ports	PCIe	Driver	XDP mode	Power	Reference build
Intel X710-DA2	2 × 10G	Gen3 ×8	i40e	native	3.3–4.5 W	Edge 1U
Intel XL710-QDA2	2 × 40G	Gen3 ×8	i40e	native	7–8 W	—
Intel E810-CQDA2	2 × 100G	Gen4 ×16	ice	native	15–21 W	Standard 1U · High-density 2U
NVIDIA ConnectX-6 Dx	2 × 100/200G	Gen4 ×16	mlx5	native	16–22 W	SmartNIC 2U

Drivers

XDP mode is what decides whether packets are processed at the driver ring or after the kernel has already built an skb. It is the single largest performance factor in the whole hardware decision, and it is a property of the driver, not of BNGSOFT.

Driver	Typical hardware	XDP mode	Notes
i40e	Intel X710 / XL710	native	Well-trodden path at 10G and 40G.
ice	Intel E800 series (E810)	native	The 100G workhorse. Both 100G reference builds use it.
mlx5	NVIDIA ConnectX-5 / 6 / 6 Dx	native	Adds hardware conntrack offload on 6 Dx.
ixgbe	Intel 82599 / X520 / X550	native	Older 10G silicon. Supported; no reference build.
bnxt_en	Broadcom NetXtreme-E	native	Supported; no reference build.
vmxnet3	VMware virtual adapter	generic only	Read this before virtualising. vmxnet3 has no native XDP path, so packets are processed after skb allocation. Expect a large fraction of bare-metal throughput to disappear. Fine for labs, demos and functional testing — not a production data plane.

The virtualisation trap. A BNG that benchmarks well on bare metal and badly in a VM is almost always a generic-XDP problem, not a BNGSOFT problem. If you are evaluating in VMware on vmxnet3, you are measuring the driver, not the platform. Test on a passthrough NIC with a native XDP driver, or treat the virtual result as a functional check only.

Reference builds

Build	NIC	CPU	RAM	Capacity	Subscribers	Power
Edge 1U	X710-DA2	8-core	32 GB	~20 Gbps	~6,700	~250 W
Standard 1U	E810-CQDA2	16-core	64 GB	~100 Gbps	~33,000	~380 W
High-density 2U	2 × E810-CQDA2	2 × 16–24-core	128 GB	~200 Gbps	~64,000	~650 W
SmartNIC 2U	ConnectX-6 Dx	24-core	128 GB	~200 Gbps	~64,000	~600 W

Memory

Populate every channel

A packet-rate-bound data plane is memory-bandwidth sensitive. A CPU with half its channels filled gives up throughput that extra capacity cannot recover.

40G class — 32–64 GB, all channels populated
100G class — 128 GB, all channels populated
State is modest: ~8 KB userspace plus BPF state per subscriber. Size for bandwidth, not tables.

Platform floor

Commodity x86, current kernel

Kernel — 5.10 minimum, current 6.x recommended
Architecture — x86-64 Intel or AMD; ARM64 in development
Bonding — 802.3ad LACP
Boot media — USB flash or SATA DOM

Bring us your numbers, not ours

Send your subscriber count and your measured busy-hour average and we will work the sizing against your traffic, then hand you a build to test on your own hardware. Core count comes out of that test — not out of a spreadsheet.

Frame-size arithmetic assumes standard Ethernet framing with 20 bytes of per-frame wire overhead (7 B preamble, 1 B SFD, 12 B inter-frame gap). The 50 Mpps engine ceiling is a measured platform figure; throughput values derived from it are arithmetic, not independent measurements. Reference build capacities and subscriber counts are as published in the BNGSOFT hardware sizing brief. Adapter prices are indicative only. XDP mode reflects upstream Linux driver capability and should be confirmed against your own kernel version.