We delivers a next-generation Ethernet switching platform designed for the evolving needs of AI, HPC, storage, and cloud infrastructure. Covering speeds from 100G to 800G and up to 1.6T, the portfolio is built to support ultra-large AI data centers, high-performance computing clusters, NVMe-oF storage networks, and hyperscale cloud environments.
100G-1.6T AI Switches Powered by Marvell Teralynx
At its core, the platform is powered by advanced merchant silicon such as Marvell Teralynx. Combined with a carefully optimized hardware architecture, our switches offer ultra-low latency, high-bandwidth connectivity, and lossless networking capabilities. Together with an enterprise-grade SONiC-based operating system, this creates an open, scalable, and high-performance Ethernet foundation—without the constraints of vendor lock-in.
The hardware itself is engineered to handle massive scale and demanding workloads. From high-density 100G and 400G leaf switches to flagship 800G and 1.6T core systems, the portfolio provides a complete range of deployment options. Efficient thermal design and power optimization ensure stable operation even in dense environments, while tight hardware-software integration allows the switches to consistently deliver line-rate performance under heavy load.
1.6T/800G|128-Port MPO 102.4 Tbps Low Latency <450ns | 800G|64-Port OSFP 51.2 Tbps Low Latency 560ns |
400G|64-Port OSFP 25.6 Tbps Low Latency 560ns |
25.6 Tbps Low Latency 560ns |
|---|---|---|---|
400G 32-Port QSFP-DD 12.8 Tbps Low Latency 500ns | 200G|64-Port QSFP56 12.8 Tbps Low Latency 500ns | 100G|64-Port QSFP28 6.4 Tbps Low Latency 500ns | 100G|32-Port QSFP28 3.2 Tbps Low Latency 500ns |
CX9128E-N(CPO Switch)
CX764QO-N
CX764QD-N
~500ns Single-Hop Latency: Overcoming Latency Bottlenecks for AI & HPC
Latency is a critical factor in AI and HPC environments, especially for distributed training where GPUs must constantly synchronize data. Asteraix addresses this with a highly optimized forwarding architecture, achieving single-hop latency of around 450- 500 nanoseconds. By minimizing queuing and processing overhead, the network effectively becomes a near-zero-latency data path, enabling faster model training and more efficient GPU-to-GPU communication.
ROCEv2 Outperforming InfiniBand in Real-World AI Workloads
In real-world AI workloads, this optimized Ethernet approach has demonstrated strong performance compared to traditional InfiniBand networks. Tail latency (P90) can be reduced by over 20%, while overall training efficiency improves significantly. This shows that open Ethernet, when properly optimized, can match—and even outperform—proprietary architectures while offering greater flexibility and scalability.
Hardware-Optimized NVMe-oF High-Performance Storage Networking
To meet the high-throughput requirements of flash storage and instantaneous AI training data reads, our switches undergo hardware-level buffer and queue scheduling optimizations:
• Demonstrates outstanding physical throughput and efficiency under standard 4KB/8KB NVMe-oF high-performance storage workloads.
• Reduces physical write latency by 10.1% compared to traditional commercial InfiniBand storage solutions.
• Boosts concurrent write throughput (IOPS) by 10.5%, making Asteraix switches the ideal physical networking platform for distributed flash storage, AI dataset loading, and HPC clusters.
Hardware-Enabled AI-Native Traffic Control & Lossless Networking
While providing powerful physical throughput, the integrated Enterprise SONiC (AsterNOS) directly unleashes advanced traffic scheduling algorithms on the ASIC forwarding plane:
• ASIC-Level Traffic Scheduling: Runs Adaptive Routing and Switching (ARS), Packet Spraying, and Flowlet Switching at the hardware level. This resolves traditional ECMP-induced link congestion where one physical link is overloaded while others remain idle.
• Turnkey Easy RoCE Hardware Protection: Templates complex PFC, ECN, and DCBX parameters directly onto the physical chip. Combined with hardware PFC Watchdogs and smart buffer optimizations, it completely prevents congestion propagation and physical deadlocks.
Physical Topology Sensing: High-Precision In-Band Telemetry (INT)
Asteraix's hardware forwarding plane supports high-precision In-band Telemetry (INT), transforming the physical network into a completely transparent "glass box":
• ASIC-Level Packet Tagging: The switching silicon directly stamps telemetry data (timestamps and path identifiers) onto every single packet, tracing the physical path and nanosecond-level latency at each node.
• Dynamic Path Optimization: Perceived packet loss and queue latency are fed back to the routing engine in seconds, dynamically rerouting traffic via Dynamic Path Selection to evade congested physical links.
Highly Reliable Cloud Data Center features
Engineered to satisfy modern multi-tenant cloud environments requiring elastic physical scalability and high-density virtualization:
• Full-Line-Rate EVPN-VXLAN: Supports symmetric IRB, Anycast Gateway, and VXLAN Cross Connect at the hardware line rate to achieve seamless Layer 2/Layer 3 virtualization and multi-tenant isolation.
• High-Availability Physical Redundancy: Supports Multi-Chassis Link Aggregation (MC-LAG) to provide active-active redundant physical links.
Hardware-BFD Offloading: Offloads Bidirectional Forwarding Detection (BFD) directly to the physical silicon. This achieves millisecond-level link fault detection and self-healing without consuming CPU resources.
Modernized Open Automation & Observability Ecosystem
Asteraix switches fully embrace SDN automation and DevOps ecosystems, breaking away from traditional "closed CLI black boxes":
• Automated Mass Deployment: Native integration with Ansible, supporting Configuration as Code to deploy and provision physical devices in minutes.
• Real-Time Hardware Visualization: Integrates seamlessly with Prometheus, Grafana, and LibreNMS to display port traffic, optical module health, physical temperatures, and overall power consumption.
• Open Interfaces: Supports standard gRPC and sFlow telemetry APIs, facilitating deep integration with existing cloud management systems.
Lowering the Learning Curve
Features a Cisco-like CLI, helping traditional network engineers transition smoothly and painlessly to SONiC open networking.
Why Choose Asteraix 100G-1.6T Hardware Switching Platform？
1. Globally Proven in 50+ Countries: Successfully deployed in large-scale AI training networks, carrier cloud networks, top-tier European internet data centers, and enterprise private clouds.
2. Hardened Enterprise SONiC: Offers superior stability, faster physical feature releases, and deep hardware-specific ASIC optimization compared to community versions.
3. Full-Lifecycle Expert Support: Delivers comprehensive services including initial physical network pre-design, on-site deployment consultation, and rapid global technical response.
4. Vendor Decoupling with Unmatched TCO: Empowers enterprises to build open, high-performance Ethernet infrastructures, drastically reducing the total cost of ownership per Gbps.
