Most BNG comparisons fail before they start.
Performance — five questions
| Ask | Why it matters | A weak answer sounds like |
|---|---|---|
| At what frame size? | Bit rate is packet rate times frame size. At 64 B, 50 Mpps is 33.6 Gbps; at 1518 B the same engine moves 615 Gbps. Without the frame size the Gbps figure cannot be checked. | “Up to 100 Gbps.” |
| Which features were enabled? | Stateless L2 forwarding and a full subscriber path are different workloads by roughly 6×. VPP’s own published figures show 115–118 cycles/packet stateless against 359 cycles/packet for CGNAT. | “Line rate.” |
| Was the subscriber table loaded? | Headline packet rates are usually measured with an empty state table. Real gateways carry tens of thousands of sessions, and that is where architectures separate. | “40+ Mpps per core.” |
| Is this observed, validated, or port capacity? | Three different numbers that routinely get quoted as one. Installed NIC bandwidth is a chassis bound, not a forwarding result. | A single number with no qualifier. |
| What was the loss threshold? | “Line rate” at 0.1% loss and at zero loss are not the same claim. | Silence. |
Architecture — four questions
| Ask | Why it matters |
|---|---|
| Where are packets processed? | Purpose-built silicon on line cards, kernel-bypass in userspace, or in-kernel at the driver. Each has real consequences for cost, upgrade behaviour and operational visibility. |
| Do standard tools still work? | If the port is claimed by a bypass framework, ip, ethtool and tcpdump stop seeing it. That changes how every incident is diagnosed at 3am. |
| How many forwarding tables exist? | A synchronised copy of the FIB is a copy that can drift. Ask what happens when it does. |
| What happens during an upgrade? | Process restart, chassis failover, or an atomic in-place swap. Ask for the subscriber impact in seconds, not adjectives. |
Commercial — three questions
| Ask | Why it matters |
|---|---|
| What exactly is metered? | Per-subscriber licensing tied to hardware behaves very differently from per-traffic licensing on commodity servers when you grow 30% in a year. |
| Is CGNAT a separate licence? | On several platforms it is. That single line moves five-year cost more than most performance differences. |
| What is the power and rack budget? | Chassis platforms run kilowatts; commodity servers run hundreds of watts. Over five years, in a colo, that is a capital-scale number. |
Where our own answer is the weaker one
Every architecture trades something. Ours trades these, and we publish them rather than waiting for you to find them in a proof of concept:
At 100 pps, kernel-bypass designs reach around 2 µs where an in-kernel path sits near 82 µs. The gap closes sharply under real load — roughly 7 µs against 3 µs at 1 Mpps — but at idle it is real.
An XDP data plane cannot hold a packet back. Congestion is handled by marking and dropping rather than by sophisticated scheduling disciplines. If your design depends on deep hierarchical queueing behaviour, weigh this carefully.
On 64-byte frames with no state, DPDK reaches about 43.5 Mpps single-core against roughly 24 Mpps for XDP. Once the subscriber table is loaded both land in single digits per core — but the stateless number does favour bypass.
We ship a tuned kernel, not a package you install onto an existing distribution. If your operations model requires running on your own standard build, that is a genuine constraint.
The detailed comparisons
Each of these is a full brief with its sources cited, not a feature grid:
| Brief | What it covers |
|---|---|
| vs Cisco ASR 9000 & Nokia 7750 SR | Chassis-and-NPU against commodity x86: subscriber density per chassis, licensing metering, power draw from vendor calculators, and L4S availability. Estimates marked illustrative are labelled as such. |
| XDP vs DPDK | In-kernel against kernel-bypass, argued from FD.io’s own published benchmarks. Eighteen dimensions, named objections answered, and the concessions above stated in full. |
| vs NetElastic & MikroTik | Software-BNG against software-BNG — the closest comparison, and the one where architecture differences matter most. |
| vs MikroTik RouterOS | Where RouterOS stops scaling for subscriber termination, and where it remains the right answer. |
Run the questions past us first
Take the list above to every vendor on your shortlist. If our answers do not hold up under it, you have lost an afternoon rather than a procurement cycle.
When you are ready, a proof of concept runs on your hardware against your traffic — because that is the only comparison that ends the argument.