AI Fabric & Data Center Switches
Enterprise SONiC NOS Based Low Latency Data Center Switches
Skyward Telecom provides robust and reliable SONiC-based solutions, meticulously tested and hardened to guarantee seamless deployment across our full portfolio of open networking data center switches. Whether you’re building AI/ML training clusters, scaling HPC environments, or architecting RoCE/NVMe-based lossless data center fabrics, Asterfusion delivers the high-performance switching foundation you need — with ultra-low latency, high throughput, and intelligent congestion control built in.
Enterprise SONiC NOS Data Center TOR Switch, 48x25Gb SFP28, 8x100Gb QSFP28
The CX308 ensures seamless integration into modern data center fabrics with enterprise-grade SONiC operating system pre-installed.
Learn More
Data Center Low Latency Leaf/Spine Switch with ROCEv2&EVPN Multi-homing, Enterprise SONiC NOS
The CX532P-N is a 32x100G QSFP28 data center switch, engineered for Top-of-Rack (ToR) and spine architectures.
Learn More
Enterprise SONiC Distribution Data Center Switch, 32 x 400G QSFP-DD
The CX732Q-N is a high-performance 32-port 400G switch designed for the super spine layer in next-generation CLOS network architectures.
Learn More
64 Port 200G QSFP56 Low Latency Leaf Spine Switch for AI/ML/Data Center
The CX664P-N switch pre-installed with Enterprise SONiC, it supports cutting-edge features including RoCEv2 and EVPN multihoming, pairing lightning-fast forwarding rates with massive packet buffers.
Learn More
64 Ports 100G QSFP28 Data Center Switch Enterprise SONiC NOS Installed
The CX564P-N 100G data center switch is pre-installed with Enterprise SONiC, it supports advanced features such as RoCEv2 and EVPN multihoming, combining lightning-fast forwarding rates with massively oversized packet buffers.
Learn More
64 Ports 800G Ultra Ethernet Switch for AI/ML/Cloud Data Center/HPC
The CX864E-N switch pre-installed with Enterprise SONiC, it packs cutting-edge features like RoCEv2 and EVPN multihoming, delivering lightning-fast forwarding and massive packet buffers for unmatched performance.
Learn MoreData Center Infrastructure Architecture
- Manages and optimizes the communication of parallel computation workloads across GPU nodes during Al/ML inference tasks, including KV caches,activations, and tensor slices.
- Connects only to inference GPU nodes, operating in complete isolation from other networks.
- Fully supports RoCE, ensuring lossless transmission of xCCL（e.g.,NVIDIA’s NCCL） and MPI traffic.
- Product portfolio: The switch can be either CX864E-N or CX732Q-N.
- Provides storage and retrieval of machine learning models, input/output data, intermediate results, and logs, enabling data access and management for GPU-based inference workloads.
- Connects GPU nodes and storage nodes.
- Fully supports RoCE, ensuring lossless transmission of GPU-Direct RDMA, Ceph, and related traffic.
- Product portfolio: CX664D-N.
- Orchestrates the distribution of user requests to inference nodes and delivers results，Supporting multimodal input and output.
- Connects to APl Gateway, Load Balancer,and GPU nodes.
- Fully supports BGP and VXLAN EVPN to realize multi-tenant isolation, leveraging ECMP and BGP multihoming for high reliability.
- Product portfolio: CX564P-N , CX532P-N , CX308P-48Y-N-V2.
