Most broadband networks still run subscriber termination and carrier-grade NAT as two separate boxes: a BNG at the edge, and a CGNAT appliance somewhere behind it. That split made sense when both were proprietary hardware. On a modern x86 server it mostly costs you an extra hop, an extra upgrade cycle and a harder time answering the question every regulator eventually asks: which subscriber had this public address and port at this time?
Why BNG and CGNAT were separate in the first place
Chassis BNGs were built to terminate PPP and apply per-subscriber policy. NAT was an afterthought, so operators bought a dedicated CGNAT appliance when IPv4 ran out and put it in front of the core. Two vendors, two licences, two places where a subscriber’s identity lives, and two maintenance windows.
What changes when both run in one data plane
- The subscriber is known at the moment of translation. The same process that terminated the PPPoE or IPoE session assigns the port block, so the NAT log is tied to the session, not reconstructed later from two systems.
- QoS sees the subscriber before NAT. Hierarchical QoS shapes each subscriber to plan rate on the private address, before many subscribers are folded behind one public IPv4.
- Logging stays small. Logging port-block allocations instead of every flow gives the same legal answer with a few thousand records a day rather than billions.
- One upgrade path. One image, one reboot when moving between major versions, then live updates that keep every session up.
What it looks like in numbers
On BNGSOFT’s XDP data plane, a single commodity server forwards 50 Mpps with PPPoE, CGNAT, hierarchical QoS and ACLs all enabled. Translated packets never enter the Linux network stack, which is why full CGNAT in XDP uses 26 to 36 percent less CPU than kernel NAT for the same traffic. A port block is never shared by two subscribers, with a 60-second quarantine before reuse, and in the OrionOS 7.05 lab run a live upgrade at 128,000 subscribers caused zero session flaps.
When to keep CGNAT separate anyway
Combining is not always right. If one central NAT pool serves many BNGs from different vendors, or if translation has to sit at a different point in the network for routing reasons, a standalone translation node still makes sense. The point is that it should be a configuration choice on the same software, not a different product. The packet path explorer shows which stages run in each deployment model.
Sizing a combined BNG and CGNAT box
Packets per second is the limit, so throughput follows frame size: 33.6 Gbps at 64-byte frames and about 144 Gbps at IMIX. Plan RAM at 32 GB minimum, 64 GB at 40G and 128 GB at 100G, with every memory channel populated. To see how many public addresses you would need, use the IPv4 savings calculator; for throughput and box count, the BNG sizing calculator.
Where to go next
- vBNG / BRAS: PPPoE and IPoE termination with RADIUS AAA and CoA.
- XDP CGNAT: port-block NAT44, NAT64 and MAP-T / MAP-E.
- The XDP data plane: why one server reaches 50 Mpps.
- Request a proof of concept on your own hardware and traffic.


