Same user, same site, completely different transport architecture — latency, routing complexity, and UPF placement compared

1. What Actually Differs — Beyond the Marketing

The marketing story around NSA vs SA is simple: NSA uses an LTE anchor, SA goes direct to 5G core. What the marketing does not explain is what this means for the transport network — specifically for the number of interfaces, the routing topology, the latency profile, and the engineering work required to move from one to the other.

In NSA (Option 3x, the dominant early deployment), the LTE eNB is the master node. The NR gNB is the secondary node. All RRC signalling goes via the LTE eNB to the EPC (Evolved Packet Core). The NR gNB carries only user plane traffic — and that user plane traffic is split at the PDCP level between LTE and NR. The transport network must support this split, which means two separate backhaul paths converging at a single SGW-U/PGW-U in the EPC.

In SA, the LTE dependency is gone. The gNB connects directly to the 5GC AMF for control plane (N2 interface) and directly to the UPF for user plane (N3 interface). Simpler conceptually. But the transport implications are significant — you now need N2 and N3 connectivity from every gNB site to the 5GC, and the UPF placement becomes a first-order design decision that directly impacts latency.

2. Real Network Architecture — NSA and SA Side by Side

NSA Architecture — Interface Map

Interface	From	To	Protocol	Transport Requirement
S1-MME	eNB (LTE master)	MME (EPC)	SCTP/IP	Low latency, reliable — N2 equivalent in EPC
S1-U	eNB (LTE master)	SGW-U (EPC)	GTP-U/UDP/IP	User plane — carries LTE PDCP split traffic
X2-U	eNB	gNB (NR secondary)	GTP-U/UDP/IP	CRITICAL — must be < 10ms RTT for PDCP split to work
X2-C	eNB	gNB (NR secondary)	SCTP/IP	Control signalling between master and secondary nodes
NR backhaul (user plane only)	gNB	SGW-U / UPF via EPC	GTP-U/UDP/IP	NR user plane — no direct 5GC connectivity

SA Architecture — Interface Map

Interface	From	To	Protocol	Transport Requirement
N2	gNB	AMF (5GC)	SCTP/IP	Control plane — replaces S1-MME, needs reliability and low latency
N3	gNB CU-UP	UPF (5GC)	GTP-U/UDP/IP	User plane — all traffic, no LTE split — UPF placement critical
F1-C	DU	CU-CP	SCTP/IP	Internal RAN split — within transport domain
F1-U	DU	CU-UP	GTP-U/UDP/IP	Internal RAN split — fronthaul/midhaul boundary
Xn-C	gNB	Neighbouring gNB	SCTP/IP	Inter-gNB control — handover, CA coordination

Key observation: In NSA, the X2 interface between eNB and gNB is latency-critical. If your eNB and gNB are on different transport rings without direct interconnect, X2 RTT can easily exceed 10ms and PDCP split performance degrades — you see increased out-of-order delivery and throughput below expectation. This is the most common NSA transport problem operators encounter.

3. Step-by-Step Real Flow — Same User, Two Architectures

NSA: User Downloads a File in Muscat

The user is on a dual-connected device (LTE + NR). The eNB at the Muttrah site is the master node. The NR gNB at the same site is the secondary. The transport path for this user’s data:

UE transmits NR data. NR gNB receives, performs PDCP processing, sends user plane data via X2-U to the eNB master node.
eNB aggregates LTE and NR user plane data. Forwards combined stream via S1-U to the SGW-U at Muscat EPC.
SGW-U forwards to PGW-U. PGW-U applies policy, assigns UE IP, routes to internet.
Return path: Internet → PGW-U → SGW-U → eNB → splits to LTE and NR bearers via X2 → UE.

The transport carries: an X2 path between eNB and gNB (same site or cross-site), and a backhaul from eNB to EPC. The gNB does NOT have a direct path to the EPC for user plane in NSA Option 3x — everything goes through the eNB master.

SA: Same User, Same Site, Standalone 5G

After migration to SA, the LTE dependency is removed. The same gNB now operates independently:

UE connects via NR only. gNB CU-CP establishes RRC via N2 to AMF. AMF authenticates, sets up PDU session, assigns UPF.
gNB CU-UP establishes GTP-U tunnel directly to UPF via N3. No X2, no eNB involvement.
User data: UE → gNB CU-UP → N3 (GTP-U) → UPF → internet. Clean, direct, two-hop transport path.
Return path: Internet → UPF → N3 (GTP-U) → gNB CU-UP → UE.

The transport carries: N2 (SCTP) from gNB to AMF, and N3 (GTP-U) from gNB CU-UP to UPF. Two interfaces, both going directly to the 5GC. Simpler routing, but now every single gNB site needs reachability to the AMF and the UPF.

4. Latency Comparison — NSA vs SA

Path Segment	NSA Latency	SA Latency	Difference
UE to eNB/gNB (air interface)	~4ms (LTE) / ~1ms (NR)	~1ms (NR only)	SA saves 1-3ms on air interface
gNB to core (backhaul)	gNB → eNB (X2: 2-5ms) + eNB → EPC (5-15ms)	gNB → UPF direct (3-8ms if UPF near RAN)	SA can save 4-12ms if UPF is co-located
Core processing	EPC (SGW+PGW): 3-5ms	5GC UPF: 1-2ms	SA 5GC is lighter — 1-3ms saving
Total one-way latency	~15-25ms typical	~5-12ms with edge UPF	SA can be 2-3x lower latency

The latency advantage of SA is not automatic. It depends critically on UPF placement. If the UPF is centralised at a remote DC (as it often is in early SA deployments for operational simplicity), the N3 path latency can be as high or higher than NSA backhaul. The real latency win in SA comes when the UPF is deployed at or near the RAN aggregation hub — and that is a deliberate architecture decision that must be made at design time.

5. Routing Complexity — What Changes

NSA: Routing Is Simpler but X2 Is the Hidden Problem

In NSA, routing complexity is lower from a core network perspective — all traffic converges at the EPC, and gNBs only need basic IP connectivity to their eNB master and to the SGW-U. But X2 routing is the operational trap. X2 must be:

Low latency: < 10ms RTT for PDCP split to function correctly
Directly reachable: X2 cannot traverse a NAT boundary or a firewall that breaks GTP-U
Available between ALL gNB-eNB pairs: if cells overlap, X2 may need to exist between multiple eNBs and gNBs

In practice, operators with many 5G sites quickly accumulate a complex mesh of X2 connections. Each pair needs a routed path, and the number of paths grows quadratically with site count. This is manageable with 10-20 sites. With 200+ sites, it becomes an operational challenge.

SA: Routing Is Simpler in Topology but Demands More from Transport

In SA, X2 is replaced by Xn (inter-gNB interface for handover coordination). Xn does not carry user plane traffic — it is signalling only and is less latency-critical than X2. The routing complexity shifts to N2 and N3:

Every gNB needs N2 reachability to at least one AMF — typically via an SCTP multi-homing configuration for redundancy
Every gNB needs N3 reachability to its assigned UPF — and UPF selection is done by the AMF/SMF, so the transport must support reachability to multiple UPFs
UPF placement decisions directly determine which transport VRFs and routes must exist at which PE routers

Field Note: In a GCC operator’s SA rollout, the transport team discovered that AMF N2 connectivity required SCTP multi-homing across two different PE routers for redundancy. This required ECMP-capable SCTP routing in the transport VRF — a configuration that most operators had not needed before. It was a three-week delay to design, test, and deploy. Plan N2 redundancy transport early.

6. Impact of UPF Placement — The Most Consequential SA Decision

UPF Placement	N3 Latency	Use Case	Operational Trade-off
Central DC (single national UPF)	15-30ms	Initial SA launch — simplicity priority	Easy to operate, poor latency for edge use cases
Regional hub (per-region UPF)	5-10ms	Consumer 5G — balances latency and cost	More UPF instances to manage, better user experience
Aggregation hub (MEC UPF)	1-3ms	Enterprise slices, gaming, AR/VR	Best latency, most complex — requires MEC platform
Co-located with CU (ultra-edge)	< 1ms	URLLC industrial — Oman free zones	Highest cost, narrowest use case — not for consumer

7. Common Issues in the Field

X2 latency exceeding 10ms in NSA — the most common NSA performance problem. Occurs when eNB and gNB are on different aggregation rings without direct interconnect. The GTP-U packets traverse multiple hops to reach the eNB. PDCP reordering buffer fills, throughput drops, and the NR addition provides less benefit than expected.
NSA to SA migration breaking existing transport configs — during migration, gNBs are reconfigured to connect directly to 5GC. N2 and N3 routes must be provisioned in the transport VRF. If transport team is not coordinated with RAN team on the cutover schedule, gNBs come up in SA mode with no reachable AMF or UPF.
UPF placed too far from RAN in initial SA — common in early SA deployments where a single UPF is deployed at the central DC. Latency is acceptable for basic broadband but fails for enterprise SLAs and MEC use cases. Architecturally expensive to fix later.
N2 SCTP multi-homing not properly tested — SCTP multi-homing for N2 redundancy relies on the AMF and transport network correctly routing SCTP packets to alternate paths when the primary fails. This is often only tested under full failure conditions — not tested in brownout or partial failure scenarios where routing may prefer the degraded path.
Xn connectivity gaps after SA migration — not all neighbouring gNB pairs have Xn configured, causing handover fallback to N2-based handover (slower). Verify Xn mesh is complete for all overlapping cell pairs.

8. Troubleshooting Approach

NSA: measure X2 RTT first — if PDCP split performance is poor, run a ping or TWAMP test between eNB and gNB IP addresses on the X2 path. If RTT > 5ms, check whether the X2 traffic is routing via an indirect path. Direct routing or X2 traffic steering via SR Policy is the fix.
SA: verify N2 reachability from every gNB — ping AMF IP from gNB management plane. Then verify SCTP session establishment in AMF logs. A gNB that can ping AMF but cannot establish SCTP is typically blocked by a firewall or has incorrect SCTP port configuration.
SA: verify N3 GTP-U tunnel state — check UPF session table for the gNB’s N3 GTP-U tunnel. A tunnel in the table but with zero packet counters indicates an MTU or routing problem on the N3 path.
Compare NSA and SA latency empirically — after SA migration, measure RTT from UE to a known server in both NSA and SA modes. If SA is not showing latency improvement, UPF placement or transport routing is not optimised.

9. Design Recommendations — Consultant Level

Design N2 redundancy transport before SA launch — SCTP multi-homing requires routes to both AMF instances to be present and reachable. Design the transport VRF to support ECMP paths to multiple AMFs from day one. Do not add this as an afterthought.
Deploy regional UPFs from launch — avoid the single central UPF trap. Even in a small country like Oman, a single UPF at Muscat DC means Salalah users experience 30-40ms additional round-trip latency. Deploy at minimum a northern (Muscat) and southern (Salalah) UPF at SA launch.
Eliminate X2 RTT problems before NSA scale-up — if you are scaling NSA to more than 50 sites, audit X2 RTT for every eNB-gNB pair. Sites with X2 RTT > 5ms need direct interconnect or traffic engineering via SR Policy to a lower-latency path.
Keep NSA and SA transport VRFs separate during migration — do not reuse the same VRF for both EPC backhaul and 5GC N3. The route tables, QoS policies, and routing designs are different. Mixing them creates troubleshooting nightmares and increases the risk of route leakage.
Plan UPF IP addressing for transport scalability — UPF N3 IP addresses must be reachable from all gNB CU-UP nodes in the transport network. Design the IP addressing scheme so UPF subnets can be advertised as aggregated prefixes in the transport VRF — not as /32 host routes that explode the routing table.

10. Summary — Practical Takeaways

NSA and SA are not just different core network architectures — they are fundamentally different transport designs. NSA requires X2 management and EPC backhaul engineering. SA requires N2/N3 VRF design, UPF placement decisions, and SCTP multi-homing. The transport team needs to understand both and plan the migration carefully.

The single most impactful decision in a 5G SA deployment is UPF placement. Get it right and SA delivers the latency improvements subscribers and enterprises expect. Get it wrong and SA is just NSA without the LTE anchor — same or worse performance with more complexity.

Takeaway	Action
X2 RTT < 5ms is critical for NSA performance	Audit X2 RTT at every eNB-gNB pair — fix with direct routing or SR Policy
SA UPF placement determines latency	Deploy regional UPFs at aggregation hubs — never start with a single central UPF
N2 SCTP multi-homing needs transport design	Design dual-AMF reachability in transport VRF before SA launch
Keep NSA and SA VRFs separate during migration	Avoid route leakage and troubleshooting complexity during transition period
Plan UPF IP addressing for route aggregation	Use summarised subnets for UPF N3 addresses — prevent routing table explosion

NSA vs SA Transport What Actually Changes in Real Deployment