OcNOS 6.6.1 is now generally available. This release delivers the complete AI fabric solution on Broadcom Tomahawk 5-based open hardware — building on the PFC and ETS foundations introduced in OcNOS 6.6 to provide a production-ready lossless RoCEv2 fabric for AI/ML GPU clusters. Service provider deployments benefit from additional Segment Routing enhancements and expanded EVPN coverage.
Complete AI Fabric on Broadcom Tomahawk 5
OcNOS 6.6.1 DC delivers full lossless AI fabric support on Tomahawk 5 platforms including the UfiSpace S9321-64EO (64×800G OSFP, 51.2 Tbps). This includes Priority-Based Flow Control (PFC), Enhanced Transmission Selection (ETS), DCBX auto-negotiation, Explicit Congestion Notification (ECN), and RDMA-aware queue management — all validated for production RoCEv2 GPU workloads.
| 기능 | OcNOS 6.6 | OcNOS 6.6.1 |
|---|---|---|
| PFC (Priority-Based Flow Control) | Initial (select platforms) | ✓ Full TH5 support |
| ETS (bandwidth allocation) | Initial | ✓ Full TH5 support |
| DCBX auto-negotiation | Partial | ✓ IEEE and CEE modes |
| ECN (congestion notification) | No | ✓ Full ECN marking |
| 800G platform (TH5) | No | ✓ UfiSpace S9321-64EO |
| RoCEv2 queue profiles | No | ✓ Pre-validated profiles |
! OcNOS 6.6.1 DC -- Complete AI fabric configuration on TH5
!
! Step 1: DSCP-to-CoS mapping for RoCEv2
qos map dscp-cos AI-FABRIC-MAP
dscp 24 cos 3 ! RoCEv2 (CS3) -> priority 3
dscp 46 cos 5 ! Expedited (EF) -> priority 5
!
! Step 2: PFC on RoCEv2 priority
interface Ethernet1/1
qos map dscp-cos AI-FABRIC-MAP
priority-flow-control mode on
priority-flow-control priority 3 no-drop ! Lossless for RoCEv2
priority-flow-control priority 5 no-drop ! Lossless for control
!
! Step 3: ETS bandwidth guarantees
qos scheduler-group AI-SCHEDULER
strict-priority 7 ! Highest: network control
wrr cos 5 weight 10 ! 10%: routing/signaling
wrr cos 3 weight 70 ! 70%: RoCEv2 (GPU all-reduce)
wrr cos 0 weight 20 ! 20%: best-effort
!
! Step 4: ECN marking threshold
qos wred AI-WRED
cos 3 min-threshold 80 max-threshold 95 drop-probability 100
ecn ! Mark ECN instead of drop
!
! Step 5: DCBX (auto-negotiate with GPU servers)
interface Ethernet1/1
dcbx enable
dcbx version ieee ! IEEE 802.1Qaz
!
! Step 6: Apply scheduler to interface
interface Ethernet1/1
qos scheduler-group AI-SCHEDULER egress
!
! Verification:
show priority-flow-control
show dcbx interface Ethernet1/1
show qos scheduler-group AI-SCHEDULERSegment Routing Enhancements (SP)
OcNOS 6.6.1 SP expands Segment Routing capabilities with additional Flex-Algo improvements including refined delay advertisement thresholds, improved algorithm convergence during topology changes, and enhanced SR-TE policy path computation. The RSVP Auto-Bypass feature is also introduced in this release — see the RSVP Auto Bypass guide for details.
EVPN Service Expansion
OcNOS 6.6.1 expands EVPN coverage with improvements to EVPN-ETREE service handling, enhanced multi-homed ESI mass withdrawal timing, and improved EVPN IRB (Integrated Routing and Bridging) gateway performance for data center environments requiring simultaneous L2 and L3 forwarding within the same VNI.
IP Infusion Product Team