/
64 x 800Gb OSFP AI Switch, 51.2Tbps, Sub-560ns Latency
$46,800.00
- 51.2 Tbps in 2RU: Supports 64×800G, 128×400G, or 512×100G breakout modes.
- Ultra-low latency: Sub-560 ns port-to-port.
- Massive buffering: 200MB+ on-chip buffer optimized for RoCE.
- Precision timing: 10 ns PTP & SyncE for AI synchronization.
- Advanced telemetry: INT for real-time packet visibility and congestion control.
- Efficient power: Max power consumption 2400W under full load (64×800G SR8).
- Open AsterNOS: Enterprise SONiC pre-installed with top-tier SAI support.
- RoCE & AI fabrics: RoCEv2, BGP-EVPN, VXLAN, MLAG, PFC, ECN, DCBX, and multicast collectives.
- Fully compatible: Works with all major GPUs and SmartNICs.
- UEC ready: Line-rate programmable for future Ultra Ethernet standards.
- 800G ZR/ZR+ optics: Metro and long-haul connectivity options.
560 ns
Ultra Low Latency
200+ MB
Packet Buffer
2400W
Max Power Consumption
↓ 5% vs. IB
Training Job Completion Time
↑ 27.5% vs. IB
Inference Token Generation Rate
Designed for AI inference and training, high-performance computing (HPC), and cloud-scale storage, the Asteraix CX864E-N delivers 51.2 Tbps switching capacity in a compact 2RU form factor with 64×800G OSFP. Built on Marvell Teralynx 10, this switch supports ultra-low latency (Sub-560 ns) and line-rate L2/L3 forwarding. Enterprise SONiC (AsterNOS) is pre-installed for RoCEv2, EVPN multi-homing, and cloud-native operation. Networking- Support for all leading collective communication libraries, including NCCL, RCCL, oneCCL, BCL, MPI, and Gloo
- Optimize collective operations (Allreduce, Allgather, All-to-all, Broadcast) with built-in multicast
- Support various network topologies, including spine-leaf, fat-tree, dragonfly, butterfly, 3D mesh/torus, hyper-x, and hypercube
- Full RoCEv2 data path with PFC, ECN, DCBX, and INT for lossless AI fabrics
- BGP-EVPN, VXLAN, and MLAG for scalable leaf-spine and multi-tenant deployments
Use case scenarios
- AI / ML training and inference clusters
- High-performance computing (HPC) fabrics
- Cloud and colocation spine-leaf networks
- High-density data center top-of-rack (ToR)
- Storage networks with RoCEv2 optimization
AsterNOS – Enterprise SONiC, Optimized for Performance
AsterNOS is Asterfusion’s enhanced SONiC distribution, pre-installed on every CX Series switch. Built for modern data centers, it combines open-source flexibility with enterprise-grade reliability, usability, and speed. Battle-tested and production-ready, AsterNOS delivers the stability and performance you need — empowering your network to run smarter, faster, and stronger.
Layer 2 Features
Layer 3 Features
Virtualization Features
Mgmt & Monitoring
Quality of Service
Security Features
Marvell® Teralynx® 10 (98TX9180)
The Teralynx 10 is Marvell’s 51.2 Tbps programmable Ethernet switch chip built on a 5nm monolithic process, with up to 51.2 Tbps throughput. The chip integrates the industry’s lowest-latency switch and buffer fabric, a line-rate programmable forwarding pipeline, and the Teralynx® Flashlight™ telemetry engine — purpose-built for cloud data center leaf/spine, AI/ML training and inference fabrics, and HPC interconnects. Open SDK and SAI APIs deliver native SONiC support with full software compatibility across the Teralynx 7 and Teralynx 10 families.
Architecture & performance
- 51.2 Tbps full-duplex switching on a 5nm monolithic die with a 512-port radix — industry’s highest single-chip port density
- 200+ MB shared packet buffer with dynamic thresholds for microburst (µBurst) absorption
- Cut-through latency as low as 560 ns at full load; sub-600 ns across all packet sizes
- 512 × 112G LR SerDes with the industry’s lowest BER; supporting 25/50/100/200/400/800 GbE port configurations
Programmable forwarding
- Line-rate Permutable Flex-Forwarding — new protocols via software upgrade, no ASIC respin required
- Comprehensive tunneling support: IP-in-IP, GRE, MPLS, VXLAN, and Geneve
- Scalable L2/L3 forwarding tables for IPv4/IPv6; congestion-aware routing for network utilization and proactive congestion prevention
QoS, RDMA & AI fabric
- Full DCB + RoCEv2 lossless Ethernet for AI/ML training, inference, and distributed storage
- Advanced QoS with rich scheduling and shaping for multi-tenant, multi-priority workloads
- PTP (IEEE 1588v2) and SyncE timing and synchronization support for latency-sensitive applications
Telemetry & openness
- Teralynx® Flashlight™ real-time telemetry engine with P4 in-band network telemetry (INT) and critical extensions
- Enables predictive analytics, fast fault isolation, and automated remediation — delivering self-aware, self-healing networks
- Open SDK + OCP SAI APIs with native SONiC platform support; fully software-compatible with Teralynx 7 for customer migration
| Low-latency AI RoCE Switch family datasheet | Low-latency AI RoCE Switches Datasheet (PDF) |
| Quick start guide for AsterNOS CLI | AsterNOS Documentation |
The dual 1+1 hot-swappable power supplies ensure continuous operation by allowing replacement without system shutdown.
Equipped with 3+1 hot-swappable fan modules, the system delivers reliable cooling performance, maintaining stability and efficiency under demanding workloads.
| CERTIFICATE |  ISO 9001 |  ISO 14001 |  ISO 45001 |  ISO 27001 |  FCC |  CE |
| SCOPE | Quality management system | Environmental management | Occupational health and safety | Information security management | US radio frequency compliance | EU health, safety and environmental compliance |
