홈 · 문서 · 기술

Technology Deep-Dives

Engineering walkthroughs of the protocols, AI fabric architectures, and load-balancing technologies inside OcNOS — RoCEv2, DLB, GLB, Ultra Ethernet, EVPN-VXLAN, segment routing and more. Written for network architects who want substance over slogans.

Filter
Underlay · Forwarding

Segment Routing — MPLS

IS-IS / OSPF as IGP, label-switched forwarding without LDP, TI-LFA fast reroute, Flex-Algorithm constraint paths.

Read deep-dive →
Underlay · IPv6

Segment Routing v6 (SRv6)

IPv6 SID encoding, micro-SID compression, end-to-end policy with no MPLS dataplane required.

Read deep-dive →
Overlay · DC + Metro

EVPN-VXLAN

BGP-signaled multi-tenant L2/L3 VPN, ESI multi-homing, symmetric/asymmetric IRB, mass-withdraw convergence.

Read deep-dive →
Routing · Internet + DC

BGP

Add-path, BGP-LU, GR/LLGR, BMP telemetry, RPKI ROV, AS-set / AS-confederation, multi-million-route scale.

Read deep-dive →
AI Fabric · Lossless

RoCEv2

PFC + ECN + DCQCN tuning for GPU collectives. Adaptive load-balancing, per-priority telemetry, sub-µs jitter.

Read deep-dive →
Silicon · 800G AI Fabric

Tomahawk 5 Switches

BCM78900 silicon — 51.2 Tbps in one 5 nm die — and the four open 800G platforms validated on OcNOS-DC: Edgecore AIS800-64D / AS9817-64D, UfiSpace S9321-64E / 64EO.

Read deep-dive →
Silicon · Deep-Buffer 400G

Tomahawk 4 Switch

BCM56996 — 25.6 Tbps Tomahawk 4 with on-package HBM deep buffer (~70 GB). The 64×400G AI-fabric / DCI switch on OcNOS-DC: Edgecore AS9736-64D.

Read deep-dive →
Silicon · 400G DC Leaf

Trident 4 Switches

BCM56780 — 12.8 Tbps DC leaf in 1RU. Two open 32×400G platforms with iPo-DWDM (400G ZR coherent) ready: Edgecore AS9726-32DB and UfiSpace S9300-32D.

Read deep-dive →
Silicon · DC Leaf Workhorse

Trident 3 Switches

Three sub-variants — TD3-X7 (3.2 Tbps · 25G/100G), TD3-X5 (1–2 Tbps · 10G/25G), TD3-X2 (120 Gbps · 1G management). Nine OcNOS-DC validated platforms.

Read deep-dive →
Silicon · SP Edge with Timing

Qumran 2A Routers

BCM88483 — 800 Gbps SP edge with deep buffer, IEEE 1588v2 PTP Class C/D, SyncE, TSN, and 400G ZR coherent ready: Edgecore AS7535-28XB and UfiSpace S9510-28DC.

Read deep-dive →
Silicon · Compact SP CSR

Qumran 2U Routers

BCM88280 — ~360 Gbps compact SP CSR / cell-site gateway with PTP/SyncE timing. Two port-mix philosophies: Edgecore AS7515-24X (uplink-heavy), UfiSpace S9510-30XC (access-heavy).

Read deep-dive →
Silicon · Carrier Core

Jericho 2C+ Routers

BCM88850 — 7.2 Tbps single-chip / 14.4 Tbps dual-chip carrier core. Internet-scale FIB (~1.2M routes), 400G ZR every port, full IP/MPLS service plane: Edgecore AS9947-36XKB, UfiSpace S9610-36D.

Read deep-dive →
AI Fabric · Adaptive

Dynamic Load Balancing (DLB)

Sub-millisecond flowlet rebinding using live ASIC queue-depth telemetry. Closes the ECMP hash-collision gap on GPU collectives.

Read deep-dive →
AI Fabric · OcNOS 7.1

Global Load Balancing (GLB)

Fabric-wide adaptive routing — end-to-end path scoring across leaf, spine, and super-spine for 1k–16k GPU clusters. Coming in OcNOS 7.1.

Read deep-dive →
AI Fabric · Next-Gen

Ultra Ethernet (UEC)

UEC 1.0 fabric profile — packet spray, multi-path RDMA, out-of-order delivery, modern congestion control. The open answer to InfiniBand.

Read deep-dive →
AI Fabric · Reference Designs

AI Fabric Topologies

Rail-only, rail-optimized, and 3-stage Clos scheduled fabrics. Sized in port counts for 256, 1k, 4k, and 16k GPU clusters on TH4 / TH5.

Read deep-dive →
AI Fabric · Congestion Control

DCQCN — RDMA Congestion Control

WRED ECN marking, CNP feedback, and quantized rate control. NCCL-class defaults out of the box; every threshold YANG-modeled for tuning.

Read deep-dive →
AI Fabric · Watchdog

PFC Deadlock Detection & Recovery

Per-port, per-priority watchdog timer that detects paused-queue cycles and auto-drains the affected queue before training jobs hang.

Read deep-dive →
DC Fabric · Active/Active

EVPN Multi-Homing (ESI-LAG)

RFC 7432 standards-based Active/Active server attachment. Designated forwarder, aliasing, mass-withdraw — no proprietary MLAG ICL.

Read deep-dive →
DC Fabric · Inter-DC

400G Coherent DCI (ZR / ZR+)

400G ZR and OpenZR+ pluggables direct in the spine. Transponder-free metro and long-haul DCI with EVPN inter-DC for tenant extension.

Read deep-dive →
AI Fabric · Decision Guide

InfiniBand vs Ethernet for AI

Workload-specific decision guide. Where modern Ethernet (RoCEv2 + DLB + UEC) closes the gap, where IB still wins, and how to pick.

Read deep-dive →
Layer-2 · Encryption

MACsec

IEEE 802.1AE line-rate encryption — AES-GCM 256, EAPoL-MKA key management, hitless rekey on 100G/400G transport.

Read deep-dive →
Observability · gNMI

스트리밍 텔레메트리

gNMI Subscribe over gRPC, OpenConfig YANG, dial-out collectors, and integrations with Telegraf, Prometheus, and Grafana.

Read deep-dive →
Timing · 5G + AI

PTP & SyncE

IEEE 1588v2 Class C/D, SyncE G.8261/G.8262, boundary clock, transparent clock, hybrid mode, sub-100ns phase accuracy.

Read deep-dive →
Transport · RSVP-TE

MPLS Traffic Engineering

RSVP-TE LSPs, CSPF path computation, Fast Reroute link/node protection, DS-TE class-types, auto-bandwidth, inter-AS TE.

Read deep-dive →
Automation · Day-0

NETCONF, YANG & ZTP

NETCONF over SSH/TLS, OpenConfig + IETF YANG, candidate datastore with commit/rollback, Zero-Touch Provisioning, Ansible & Salt.

Read deep-dive →
No topics match this filter yet — pick a different category, or suggest a topic.

Want a topic covered?

Tell us which protocol, architecture, or trade-off you'd like a deep-dive on — we write what operators ask for.

Suggest a topic →