The Spectrum-4 is the switching ASIC at the centre of every Spectrum-X deployment. At 51.2 Tbps aggregate switching capacity it supports two port breakout configurations: 128 ports at 400 GbE each (QSFP112 cages), or 64 ports at 800 GbE (OSFP cages). For AI cluster leaf switches the 128×400GbE configuration dominates because each DGX H100 NIC runs at 400 GbE and you want one NIC per port for full non-blocking bisection bandwidth.

Forwarding latency in cut-through mode is under 300 nanoseconds for a 64-byte cell. Store-and-forward mode — which some data centre switches use to eliminate corrupt frame propagation — would take 1 to 5 microseconds for a 9,000-byte jumbo frame. For GPU collective synchronisation barriers where thousands of small acknowledgement packets must complete before the next compute phase begins, that additional 1–5 µs per hop compounds across multi-hop paths and adds measurable collective completion time. Cut-through is non-negotiable in AI fabrics.

The 48 MB shared packet buffer per Spectrum-4 chip is the single most operationally significant number in this chapter. For comparison, Broadcom Tomahawk 4 carries 12 MB of shared buffer. Four times the buffer means Spectrum-4 can absorb incast bursts at synchronisation barriers before ECN threshold is reached, giving DCQCN time to ramp flows back down before tail drops occur. This is not a theoretical advantage — in measured All-Reduce benchmarks the difference in tail latency between 12 MB and 48 MB buffer depth at the leaf is measurable at cluster scales above 256 GPUs.

The hardware adaptive routing engine is the mechanism that eliminates ECMP hash collisions. Every packet arriving at a Spectrum-4 egress port triggers a hardware lookup against the current queue depth of all candidate next-hop paths. The engine selects the least-loaded path. This decision happens at line rate — no CPU interrupt, no software polling loop. The update interval for queue depth readings is per-packet: every new packet arrival queries fresh congestion state. Compare this to ECMP, which commits a flow to a path at the first packet based on a hash of the 5-tuple and never revisits that decision even if the chosen spine link becomes saturated.

What Just Happened (WJH) is the hardware telemetry engine embedded in every Spectrum-4. When a packet is dropped — tail drop, ECN mark, TTL expiry, ACL deny — the WJH engine captures a drop event record containing the source/destination MACs and IPs, the VLAN and DSCP, the drop reason code, and a nanosecond-resolution timestamp. These records are written into a ring buffer in ASIC memory. They are accessible in two ways: via a DOCA application running on the BF3 DPU, or via a Docker container running the NetQ WJH agent directly on the switch CPU. Both paths expose the same event stream.

# Query WJH drop events directly from Cumulus (last 60 seconds)
leaf-01# docker exec wjh-container wjh_dump --layer L3 --period 60
Timestamp              SrcIP          DstIP          DSCP  Reason
2024-03-15T14:22:01.0  192.168.10.45  192.168.10.87  26    Tail drop - egress queue
2024-03-15T14:22:01.1  192.168.10.46  192.168.10.88  26    ECN marked - threshold
2024-03-15T14:22:01.3  192.168.10.45  192.168.10.90  26    Tail drop - egress queue
Total drops in period: 3 L3, 0 L2, 0 tunnel

# Compare Spectrum-4 buffer depth versus Tomahawk 4
leaf-01# nv show system resource
Resource          Used     Available  Total
Shared buffer     12.3MB   35.7MB     48.0MB    ← Spectrum-4
# Tomahawk 4 would show 12.0MB total

# Check adaptive routing engine statistics
leaf-01# nv show qos adaptive-routing statistics
Rebalanced flows:    1,247,832
Rerouted packets:      892,441
Queue depth polls: 4,891,002,211
Avg path change ms:       0.0003