RIP died partly because it measured every link the same. EIGRP's composite metric was the answer: a formula over up to five link characteristics — bandwidth, delay, reliability, load and MTU — weighted by five constants, K1 through K5. In practice, and by default, only bandwidth (K1=1) and delay (K3=1) participate; reliability and load are snapshots that would make the metric unstable if they steered routing, so K2, K4 and K5 default to zero and almost nobody changes them.
The default formula
With default K-values the classic metric reduces to: metric = 256 × ( (10,000,000 / minimum bandwidth in kbps) + (sum of delays in tens of microseconds) ).
Two details in that sentence do all the damage in exams and in real designs. Bandwidth is the minimum along the path — the bottleneck link — not a sum; delay is the cumulative total of every outgoing interface along the path. So a path's metric changes when any hop's delay changes, but only when the bottleneck's bandwidth changes.
Worked example
Take a path crossing two links: a 100 Mbps FastEthernet segment (default delay 100 µs) and a 1544 kbps T1 serial link (default delay 20,000 µs). Minimum bandwidth is 1544 kbps; total delay is 20,100 µs, which is 2010 tens-of-microseconds.
bandwidth term = 10,000,000 / 1544 = 6476 (integer division)delay term = 20,100 / 10 = 2010metric = 256 × (6476 + 2010) = 256 × 8486 = 2,172,416
Where the inputs come from
Both inputs are interface properties, not measurements. Bandwidth is whatever the interface reports or the bandwidth command asserts; delay is a fixed per-interface-type default (10 µs for GigabitEthernet, 100 µs for FastEthernet, 20,000 µs for serial) or whatever the delay command sets. This is why the standing advice is: to influence EIGRP path selection, adjust delay, not bandwidth — the bandwidth value is also consumed by QoS, SNMP graphs and other subsystems, while delay exists almost solely for EIGRP.
R1(config)# interface gi0/1R1(config-if)# delay 200 ! tens of µs — this is 2000 µsR1# show interface gi0/1 | include DLY MTU 1500 bytes, BW 1000000 Kbit/sec, DLY 2000 usec
Wide metrics
The classic formula breaks above 10 Gbps: 10⁷/10,000,000 kbps = 1, the same as at 10 Gbps, so faster links become indistinguishable. Named-mode EIGRP therefore uses 64-bit wide metrics, scaling bandwidth from a 10⁵ Mbps reference and measuring delay in picoseconds, then dividing by a configurable rib-scale factor when installing routes. Wide and classic routers interoperate — metrics are converted — but mixed networks can see surprising tie-breaks, one more reason named mode everywhere is the clean answer.
K-values are part of the adjacency contract: mismatched Ks tear down neighbourships, because two routers weighing paths differently cannot safely agree on loop-free routes. Verify with show ip protocols, which prints the active K-values and, in named mode, confirms wide metrics are in use.