High-Performance XDP BNG · CGNAT · QoS · Low-Latency · Commodity-Hardware Economics
Competitive Comparison · BNG / BRAS Platforms · with sourced specifications

Stop Paying for Chassis Silicon You Don't Need

BNGSOFT vs Cisco ASR 9000 vs Nokia 7750 SR — a real, sourced comparison. The big-iron BNGs are superb engineering built to terminate hundreds of thousands of subscribers in a single system. You pay for that ceiling in purpose-built silicon, per-subscriber licences, multi-kilowatt power and lock-in — whether you use it or not. BNGSOFT delivers chassis-class subscriber density per node on a commodity x86 server, with flexible licensing (per-traffic or per-subscriber) and no proprietary-hardware premium — and you grow by adding cheap boxes instead of fork-lifting a chassis. Figures below are cited; see Sources.
Cisco prices its BNG at one licence unit per 8,000 subscribers, on top of the chassis. Nokia bills per-subscriber on FP-silicon line cards. BNGSOFT is software on a commodity server, licensed the way that fits you.
8,000
subscribers per Cisco BNG
licence unit (à la carte)¹
Your way
BNGSOFT licence: per-traffic OR
per-subscriber — software-only
~7–26 kW
typical loaded ASR 9000
chassis power² · vs ~hundreds W/server
Scale-OUT
add cheap boxes to fit growth
— no line cards, no fork-lifts

This is a straight comparison for architects and CFOs, with vendor-published numbers cited. We name what the chassis platforms genuinely do better — then show why, for the vast majority of real BNG deployments, paying chassis economics is buying a locomotive to run a delivery van. The data plane that makes BNGSOFT fast runs in XDP at the NIC driver, before the kernel stack, so a commodity dual-socket server reaches subscriber densities that used to demand purpose-built silicon.

1 · Where the packet is actually decided

All three forward in a fast path. The difference is where and on what: BNGSOFT decides the packet in software, at the NIC driver, before the Linux kernel stack, on a commodity server; the chassis platforms decide it in purpose-built NPU/FP silicon on line cards inside a chassis.

BNGSOFT
Subscriber packet→ NIC driver · XDP ⚡ — forward + CGNAT + firewall + QoS + L4S/AQM→ Wire Linux kernel stack (bypassed)
Cisco ASR 9000
Subscriber packet→ Line-card NPU silicon · Lightspeed / Tomahawk (IOS XR)→ Wire
Nokia 7750 SR
Subscriber packet→ Line-card silicon · FP4 / FP5 (SR OS)→ Wire
The honest takeaway: at broadband scale the per-packet speed of all three is effectively equivalent — what differs is economics. BNGSOFT removes the purpose-built silicon, the chassis, the redundant supervisors and the per-subscriber licence book, and replaces them with a server you already know how to buy.

2 · The numbers, side by side SOURCED

Dimension BNGSOFT XDP BNG Cisco ASR 9000 Nokia 7750 SR
Data plane Software XDP at the NIC driver, before the kernel, commodity x86 Purpose-built NPU line cards — Lightspeed / Tomahawk (IOS XR) Purpose-built FP4 / FP5 silicon line cards (SR OS)
Per-node subscriber scale ~64k/box (PCIe-bus-bound) — scale-out 32k sessions / NPU; up to 256k+ / chassis (ASR 9922, Tomahawk)¹ Very high — multi-100k subscribers / chassis (ESM)⁵
System forwarding capacity Bus-bound: ~tens of Gbps per direction / box Up to 48 Tbps full-duplex / chassis³ Up to 216 Tbps FD / chassis (FP5)⁴
Licensing model Flexible — per-traffic OR per-subscriber (your choice); software-only, decoupled from hardware À la carte BNG licence — 1 unit = max 8,000 subscribers¹, on top of the chassis Per-subscriber / feature licensing, tied to vendor hardware
System power (typical) ~Few hundred W, 1–2U server ~7.4 kW (10-slot) to ~26 kW (16-LC) loaded chassis² Multi-kW chassis (FP5 silicon = 0.1 W/Gbit, FP4 = 0.4)⁴
L4S / low-latency AQM Built into the data plane, in production today Low-Latency DOCSIS (PIE AQM) on the cBR cable platform; ASR 9000 BNG roadmap-dependent⁶ Nokia Bell Labs pioneered L4S (RFC 9330/9331)⁷ — emerging on the SR BNG data path
CGNAT / NAT44 Built into the XDP data plane — no extra card/licence Typically a separate service card (ISM) / licence Separate ISA / ESA card or licence
Hitless upgrade Software ISSU — no redundant unit needed ISSU / NSR — needs redundant RSP hardware ISSU — needs redundant CPM hardware
Hardware unit One 1–2U dual-socket server + standard NICs Multi-slot chassis + line cards + dual RSPs Multi-slot chassis + line cards + dual CPM/SFM
Scaling model Scale-OUT: add cheap boxes to match demand Scale-UP: add line cards → bigger chassis Scale-UP: add cards → bigger chassis
Lock-in / maturity Commodity x86, no silicon lock-in; newer, focused, fast-moving Cisco silicon + licensing; decades, vast install base & support Nokia silicon + licensing; decades, vast install base & support

Green = the economically decisive win for most operators. Superscripts cite vendor / public sources listed under Sources. We hand the chassis two rows plainly: absolute single-system capacity and maturity.

3 · Subscriber density & the economics that follow

A loaded chassis dwarfs one BNGSOFT box on absolute scale — that's real, and we show it to scale. The point is what that ceiling costs, and that most PoPs never use it.

Per-system subscriber ceiling — and where a real PoP actually sits

Vendor-published maxima vs a typical regional PoP load. The chassis ceiling is huge — and you buy and power it whether or not you fill it. SOURCED + ILLUSTRATIVE
Cisco ASR 9922 (max)¹
256k+ subscribers / chassis
256k+
Nokia 7750 SR (very high)⁵
multi-100k / chassis
100k s+
Typical regional PoP load
~30k subs
what a PoP uses
BNGSOFT per 2×100G box
~64k subs
one box covers it →
Scale-out beats stranded scale-up. If a PoP needs 30k subscribers, that's a single BNGSOFT box with headroom — or two for N+1 — not a six-figure chassis provisioned for 256k "just in case," drawing kilowatts and metering you per subscriber the whole time. You add capacity the day you need it, one cheap server at a time.

4 · Latency — the honest version

There are two latencies, and vendors love to blur them.

① Raw forwarding latency
Effectively a tie — and imperceptible
  • Chassis NPU/FP ASICs forward in sub-µs to low-µs.
  • BNGSOFT's XDP path adds microseconds — ~13 ns/packet of overhead, <10 µs at typical load, ~82 µs at 1 Mpps.⁸
  • At broadband scale the last-mile RTT (milliseconds) dwarfs all of it. We won't pretend µs matter here.
② Latency subscribers actually feel — queue latency under load
This is the real game: bufferbloat vs L4S/AQM
  • Under congestion, classic deep buffers add tens to hundreds of milliseconds of queuing delay — the lag and jitter that wreck gaming, video calls and cloud apps.
  • A modern L4S / AQM data path holds working latency in low single-digit milliseconds even at saturation.
  • BNGSOFT ships L4S/AQM built into the data plane, in production today — no extra card, no separate licence.
  • Credit where due: Nokia Bell Labs pioneered L4S (RFC 9330/9331)⁷ — outstanding research. The question for the chassis is when it ships as a licensed BNG feature on your platform and release; on BNGSOFT it's included and running now.

Working latency vs offered load — bufferbloat vs L4S/AQM

The latency a subscriber experiences as the link approaches saturation. Lower and flatter is better. ILLUSTRATIVE
50 100 150 200 250 0 working latency (ms) offered load  →  idle ............................................. saturated ~200 ms+ lag & jitter flat — low single-digit ms
Classic deep buffers (bufferbloat) BNGSOFT L4S / AQM data path

5 · The TCO knockout — paying for a ceiling you'll never hit per box

Strip it to money. A chassis prices in a multi-hundred-thousand-subscriber, multi-terabit ceiling, vendor silicon, redundant supervisors, multi-kilowatt power and a per-subscriber licence meter (Cisco: one unit per 8,000 subs¹). Most PoPs use a fraction of that. BNGSOFT's unit of capacity is a cheap server you size to demand and license the way that fits.

Directional cost & power to serve a mid-size PoP — qualitative ILLUSTRATIVE

Not a quote: relative total cost (hardware capex + per-sub licences + power + sparing) and system power for a representative regional BNG. Shorter is cheaper / lower.
Chassis BNG — total cost
chassis + line cards + per-sub licences + multi-kW power + specialised spares
highest
BNGSOFT — total cost
commodity server + flexible licence + low power
a fraction
Chassis system power²
~7–26 kW loaded
kilowatts
BNGSOFT system power
~hundreds W
~0.3–0.5 kW
Licensing that fits your business — with no hardware premium underneath. Choose per-traffic (capacity-based, so subscriber growth doesn't inflate your licence) or per-subscriber — whichever matches how you sell. Either model runs on commodity hardware you own, with no proprietary chassis, line cards or NPU silicon to buy and no vendor-hardware lock-in.

6 · Where each actually fits

Pick the chassis when…
  • You must terminate hundreds of thousands of subscribers in a single system in one mega-PoP.
  • You need multi-terabit single-node throughput with in-chassis hardware redundancy.
  • You're standardised on that vendor's ecosystem with existing support and operations.
Pick BNGSOFT when… (almost everywhere else)
  • Chassis-class per-node density at a fraction of the cost, on hardware you already buy and refresh.
  • Distributed / edge BNG across many PoPs — scale each with cheap boxes, not fork-lifts.
  • You're a growing or cost-sensitive ISP and want flexible licensing (per-traffic or per-subscriber) without a proprietary-hardware premium.
  • You want modern low-latency (L4S/AQM), built-in CGNAT, anti-spoof and per-sub QoS without buying extra cards or licences.
  • You want lower power, smaller footprint, no silicon lock-in, and zero-downtime upgrades without a second redundant chassis.
Why BNGSOFT wins the deal
$
Flexible licensing, your wayPer-traffic or per-subscriber — whichever fits how you sell — vs a chassis metered at one licence unit per 8,000 subs, on top of the hardware.
↓
CapEx collapses to a serverA commodity 1–2U box replaces a chassis + line cards + redundant supervisors.
⚡
Hundreds of watts, not kilowatts~0.3–0.5 kW per node vs a typical ~7–26 kW loaded chassis — lower power and rack, per PoP.
＋
Scale-out, not fork-liftGrow PoP by PoP with cheap boxes sized to real demand — no stranded chassis capacity.
★
Low-latency built inL4S/AQM, CGNAT, anti-spoof and per-sub QoS included in the data plane — no extra cards or per-feature licences.
⛓
No silicon lock-inCommodity x86 + standard NICs. Buy, rack and refresh on the cycle you already run.

7 · Operations & management — NOC2 vs the chassis OSS stack

The chassis platforms have powerful management — but it lives in separate, heavyweight, separately-licensed OSS suites: Cisco Crosswork / EPN-Manager / NSO, Nokia NSP. They're built to manage huge multi-vendor networks and demand their own servers, integration projects and licences. BNGSOFT includes NOC2 — a lightweight, ISP-subscriber-focused operations console — with the platform, no extra OSS licence. You get chassis-class density and chassis-class operations, without the OSS price tag.

OperationsCisco / Nokia OSS (Crosswork·EPNM·NSO / NSP)BNGSOFT NOC2
PackagingSeparate products, per-feature / per-device licensing, dedicated serversIncluded with the BNG license; one web app on your infra
DeploymentCluster + integration project (weeks–months)Single binary + DB; live in hours
Subscriber triageVia OSS workflows / per-platform CLIOne search box (user / IP / MAC) → live session + history, fleet-wide
CGNAT visibilityAdd-on logging / collector toolingBuilt-in public-IP:port lookup, IPFIX, pool & exhaustion
Customer-facingBuild it yourself on top of the OSSBranded status page, outage email, "is my line up?" check — included
Alerting & healingPowerful, but configured in the OSS suitePer-customer/region alerts, capacity/anomaly, self-healing rules
NorthboundgNMI / NETCONF / OpenConfig (a genuine chassis strength)gNMI / OpenConfig read-only telemetry for any NMS
Access & auditEnterprise IAM / TACACS+RBAC + 2FA + region scoping + full audit + SSH PIN
FitBuilt for huge multi-vendor carrier networksBuilt for the ISP BNG operator — focused, fast, included
Fair framing. For a Tier-1 multi-vendor backbone, a full OSS suite (Crosswork/NSP) does things NOC2 doesn't aim to — cross-domain orchestration, MPLS/optical, thousands of nodes. NOC2's scope is the BNG/CGNAT subscriber edge, where it delivers the operations an ISP actually runs every day — triage, CGNAT, capacity, customer status — out of the box, at no extra licence. The chassis gives you a data-plane ceiling you pay for; NOC2 closes the usual gap by giving you the operations layer for free.

The bottom line

The Cisco ASR 9000 and Nokia 7750 SR are excellent machines built to terminate hundreds of thousands of subscribers in one system — and they bill you for that ceiling in silicon, per-subscriber licences (Cisco: one unit per 8,000 subs), multi-kilowatt power and lock-in, per PoP, forever. Almost no operator needs that ceiling per box.

BNGSOFT runs the entire BNG — forwarding, CGNAT, firewall, per-sub QoS, L4S/AQM low latency — in XDP before the kernel on a commodity server, delivering chassis-class per-node density with flexible licensing (per-traffic or per-subscriber), a fraction of the capex and power, no proprietary line cards, no fork-lifts and no silicon lock-in. Buy a server, not a chassis — license it the way that fits, and grow by adding cheap boxes.

Sources

  1. ¹ Cisco — Broadband Network Gateway Configuration Guide for ASR 9000 (subscriber scale 32k/NPU, 256k+/chassis; BNG licence à la carte, 1 unit = max 8,000 subscribers): cisco.com/c/en/us/td/docs/routers/asr9000/software/711x/bng/configuration/guide/ & community.cisco.com BNG scale threads.
  2. ² Cisco — ASR 9000 power specifications / Power Calculator (3-kW AC modules; example 10-slot ≈ 7,440 W; 16-line-card config ≈ 26,325 W typical / 31,520 W max; typical draw ~30% below max): cisco.com ASR 9000 hardware overview & reference guide.
  3. ³ Cisco — ASR 9000 line-card data sheets (24-port = 200 G, 48-port = 400 G; up to 48 Tbps full-duplex/chassis): cisco.com/c/en/us/products/collateral/routers/asr-9000-series-aggregation-services-routers/.
  4. ⁴ Nokia — FP5 network processor (up to 1.6 Tb/s per NP; 7750 SR-s system 2.4 Tb/s → 216 Tb/s FD; FP5 = 0.1 W/Gbit vs FP4 = 0.4 W/Gbit): nokia.com/networks/technologies/fp5/ & gazettabyte.com FP5 chipset.
  5. ⁵ Nokia — 7750 SR Enhanced Subscriber Management (ESM) / BNG documentation (very high per-system subscriber scale): documentation.nokia.com 7750 SR subscriber-management & nokia.com 7750 Service Router.
  6. ⁶ Cisco — Low Latency DOCSIS (PIE AQM, L4S) on cBR Converged Broadband Routers: cisco.com/c/en/us/td/docs/cable/cbr/ DOCSIS software configuration guide.
  7. ⁷ Nokia Bell Labs / IETF — L4S architecture & ECN protocol, RFC 9330 / RFC 9331 (pioneered by Nokia Bell Labs with BT and Simula): nokia.com/bell-labs/research/l4s/ & rfc-editor.org/rfc/rfc9331.
  8. ⁸ XDP forwarding latency — "Understanding Delays in AF_XDP-based Applications" (arXiv 2402.10513: <10 µs at low rate, ~82 µs at 1 Mpps, ~13 ns/packet overhead) & Red Hat "high-performance, low-latency networking with XDP": arxiv.org/abs/2402.10513 & developers.redhat.com.
BNGSOFT · High-Performance XDP BNG · CGNAT · QoS · L4S · Commodity-Hardware Economics
Ask us to model your fleet's TCO against your incumbent chassis quote and per-subscriber licence terms.
[ Sales / contact placeholder ]
sales@bngsoft.com · bngsoft.com
Methodology and honest framing. Competitive positioning document; product and company names — Cisco, ASR 9000, IOS XR, Nokia, 7750 SR, SR OS, FP4/FP5 — are trademarks of their respective owners, used for identification and comparison only. Figures attributed to Cisco and Nokia are taken from publicly available vendor documentation and reputable industry reporting (see Sources) and vary by chassis model, line card, software release and commercial agreement; they are not drawn from a specific customer quote. Subscriber-scale and throughput maxima are vendor-published system maxima, not guarantees for any given configuration. The Cisco BNG licence figure ("1 unit = max 8,000 subscribers") and power examples reflect Cisco's published BNG configuration guide and power-calculator documentation respectively; typical power depends on installed line cards, route processors and fabric. Items marked ILLUSTRATIVE (the latency-vs-load chart, the directional cost chart, "typical PoP load") are directional representations for a mid-size deployment, not contractual numbers — validate against your own quote and traffic. BNGSOFT figures: per-node subscriber capacity is throughput-driven — a dual-socket commodity x86 server with 2×100G NICs (PCIe gen4 ×16) can carry on the order of ≈64,000 concurrent subscribers at a ~3 Mbps busy-hour rate (capacity ≈ NIC usable line rate ÷ busy-hour per-subscriber rate, capped by a ~131,072-entry per-node table ceiling), the practical ceiling being PCIe/bus throughput with CPU headroom remaining; exact capacity depends on traffic mix, packet size, NIC model and PCIe generation. Latency: raw XDP forwarding latency figures are from the cited independent study and are workload-dependent; the working-latency (bufferbloat vs L4S/AQM) comparison is qualitative and reflects the behaviour of deep-buffer queuing versus modern AQM, not a single measured benchmark across these specific platforms. L4S credit: L4S was pioneered by Nokia Bell Labs; the comparison reflects that on BNGSOFT L4S/AQM is included in the shipping data plane, whereas on the chassis platforms low-latency/L4S availability is release-, platform- and licence-dependent and may be fully supported. BNGSOFT is commercially licensed software — not free; the operator chooses per-traffic (throughput/capacity-based) or per-subscriber licensing, decoupled from hardware. Scale honesty: a single ASR 9000 or 7750 SR chassis can terminate far more subscribers and forward far more traffic than one BNGSOFT node; BNGSOFT is a scale-out architecture (add nodes) versus the chassis scale-up model, and the comparison argues economics and fit, not absolute single-system maximums. Prepared as a management and architecture overview for large-scale operators.