The device responsible for forwarding traffic between networks is a router. A router receives a packet on one network interface, looks up the packet's destination address in a forwarding table, and retransmits it out a different interface toward another network. That store-and-forward decision, repeated hop by hop, is what connects separate Internet Protocol (IP) subnets into an internetwork.
This distinction matters because home "routers" bundle several functions, and networks carry hidden routers at every scale. What follows explains the actual forwarding mechanism, where it breaks, and how to reason about it when traffic misbehaves.
The forwarding mechanism inside a router
A router's core job is a table lookup followed by a rewrite. When a frame arrives, the router strips the Layer 2 framing, reads the destination IP address from the Layer 3 header, and searches its forwarding table for the longest matching prefix. If it finds a match, it decrements the packet's Time to Live (TTL) field, recomputes the header checksum, attaches new Layer 2 framing for the next hop, and transmits the packet. This entire sequence typically executes in dedicated hardware at line rate.
The "why" is the key insight. A host only forwards directly to destinations it believes are on its own subnet. For anything else, the host sends the packet to its default gateway and lets that device decide. Routers exist precisely because no single device holds a map of the whole internet. Each router knows only its local topology and a next hop, and the packet's destination address carries the rest of the instruction. IPv6 works identically, except routers never fragment and rely on the source to size packets correctly.
Forwarding tables come from two sources. Static routes are entered manually and never change. Dynamic routes are learned through routing protocols such as OSPF (interior) or BGP (between autonomous systems). Forwarding is the fast, per-packet lookup; routing is the slower process that populates the table. That separation is deliberate, because a lookup must finish in nanoseconds while route convergence can take seconds.
Most routers also perform network address translation (NAT) at the edge of a private network. NAT rewrites source addresses so many internal hosts share one public address, which adds state to the forwarding path. Consumer equipment handles this in software on a general-purpose CPU, while carrier and data-center routers use application-specific integrated circuits (ASICs) that sustain hundreds of gigabits or multiple terabits per second.
Quantified context: how much a router actually moves
Forwarding capacity is measured as packets per second (pps), not just bits per second, because small packets stress a router far more than large ones. At 1 Gbps, a stream of 64-byte packets arrives at roughly 1.49 million packets per second, while 1,500-byte packets arrive at about 81,000 per second. A device rated for high throughput but low pps will drop traffic under a flood of small packets.
Latency through a single router is small: forwarding delay commonly falls in the low microseconds for hardware-based routers and tens to hundreds of microseconds for software-based ones. The real cost is the number of hops. Each hop adds a TTL decrement, some queuing delay, and a forwarding decision. A path across the public internet often traverses 10 to 20 routers, and congestion at any one of them can dominate total delay.
Table scale separates device classes sharply. Consumer routers hold a few thousand routes. Enterprise routers commonly hold hundreds of thousands to about a million IPv4 routes. Core internet routers must hold the full global BGP table, which crossed roughly one million IPv4 prefixes around 2022 and has continued growing, alongside a smaller but rapidly expanding IPv6 table. A device that cannot hold the full table will silently drop routes and black-hole traffic to those destinations.
Throughput numbers are frequently overstated. A router advertised as 1 Gbps may forward only a few hundred megabits per second with NAT, firewall rules, and quality-of-service (QoS) enabled, because those features run in software. The honest measure is throughput with features on, sometimes called "routing throughput" rather than "switching capacity."
Edge cases where the router is not the forwarder
The clean answer, "the router," assumes Layer 3 forwarding between IP subnets. Real networks bend that assumption in predictable ways.
Layer 2 switches forward frames within a virtual LAN (VLAN) without any router involved. They learn MAC addresses and forward by hardware table, and they do not examine IP addresses. The moment traffic crosses between VLANs, a router or Layer 3 switch takes over.
Layer 3 switches perform the same IP forwarding as routers, just optimized for high port density inside a building or data center. Functionally they are routers; the label is marketing and form factor, not protocol.
Firewalls and load balancers often sit inline and forward packets too, but they apply policy first. A firewall may forward only 30 to 40 percent of its rated raw throughput once deep packet inspection is enabled. In security-sensitive paths, the firewall is the real forwarding bottleneck even though a router is present.
Wireless access points bridge traffic between clients and the wired network at Layer 2. A mesh system's satellite node forwards frames to the main node, and the main node's router function handles the actual inter-network hop. Confusing the mesh node with the router is a common source of blame when the wrong box is rebooted.
Software-defined and cloud networks abstract the physical router entirely. Traffic between cloud subnets is forwarded by a virtual router or fabric that the customer never sees, and the effective forwarding table lives in a control plane far from the packet's path.
Traffic fails only between subnets, never within one: suspect the router or its routes. Traffic fails within a single subnet: suspect the switch, cabling, or address configuration. Handoff traffic at a firewall or VPN edge performs poorly: the firewall's inspection engine is likely the limit, not the router. Cloud or virtualized environments behave oddly after a change: the virtual routing fabric or security group rules are involved.
Misconceptions about forwarding devices
The most persistent misconception is that the home box is one device. A typical home gateway is a router, a switch, a wireless access point, a NAT translator, a DHCP server, and often a firewall crammed into one enclosure. When someone says "the router is down," they may mean any of those functions failed. Splitting the roles, with a separate access point and switch, makes diagnosis far easier.
Another misconception is that a router sends traffic directly to the final destination. It does not. It forwards to the next hop and trusts that hop to make its own correct decision. The destination is only reached through the cumulative cooperation of every router along the path. This is why a single misconfigured router in the middle of a path can break end-to-end connectivity even when both endpoints are healthy.
Some people believe switches are faster than routers in a way that makes routers unnecessary. Switches are faster per port because they do less, but they cannot connect separate IP subnets, and their broadcast domains stay isolated. Modern Layer 3 switches blur the line, but the forwarding function is the same.
A counterintuitive professional caveat: a router can forward traffic perfectly while the network still fails. If the forwarding table points to a next hop that has gone dark, the router keeps forwarding into a black hole with no error returned. Silent failure of this kind is why operators monitor route liveness, not just interface status. Interface counters can show clean traffic while packets vanish downstream.
Practical actions: diagnose forwarding with the right device
Start by confirming which layer is failing. Run a ping to a host on the same subnet and then to a host on another subnet. Success locally plus failure remotely points to the router, its default gateway configuration, or its routes.
On the sending host, verify the default gateway address matches the router interface on that subnet. A wrong gateway makes every off-subnet packet fail immediately. Run traceroute toward the destination. The first hop should be your router. If the first hop is something else entirely, another device is intercepting or the gateway is misconfigured. On the router, check that the destination subnet appears in the forwarding table. A missing route means the packet is dropped; some routers return an ICMP "destination unreachable," many do not. Test throughput with and without features enabled. If forwarded speed collapses when NAT, firewall, or QoS is active, the router's software path is the limit, not the link. If traffic crosses VLANs on a managed switch, confirm inter-VLAN routing is configured. Without a Layer 3 interface, the switch forwards nothing between VLANs. In cloud environments, check route tables and security group rules together. A permissive route with a restrictive security group fails just as hard as a missing route.
Reconcile competing numbers before replacing hardware. A router whose rated throughput is 1 Gbps but whose NAT-plus-firewall throughput is 350 Mbps will bottleneck a 500 Mbps connection. Size for the feature-enabled number, not the headline number, and add margin for growth.
Limitations and what the answer cannot cover
The "router forwards traffic between networks" answer assumes IP and a conventional topology. On non-IP networks, such as some industrial or legacy protocols, a bridge or gateway performs the equivalent function. On encrypted overlays, such as VPNs or service meshes, the physical router forwards ciphertext while a virtual device decides the inner routing, so the responsible device depends on which layer you ask about.
Forwarding also depends on the control plane being healthy. A router with a stale or incomplete table forwards incorrectly no matter how fast its hardware. This is why operators treat route convergence and table completeness as first-class health signals alongside link status and throughput.
Finally, no single device is responsible end to end. Each router along the path forwards independently, and end-to-end success requires every hop to be configured correctly. The router is the unit of inter-network forwarding, but the path is the real system. Diagnosing forwarding problems means checking each hop, not just the first one.