MetaX Xijing S600 Super Node: 64-GPU Cabinet, Zero-Cable OEX Architecture & MetaXLink Interconnect

In July 2025, MetaX (沐曦股份) announced a family of AI super-node form factors built around its new C600-series GPUs. The flagship C600 chip — MetaX's first fully domestic design-to-package supply chain — delivers performance positioned between NVIDIA's A100 and H100 and integrates the MetaXLink super-node extension fabric.

Four AI Compute Node Form Factors

Xiyun C500X Optical-Interconnect 64× GPU Super Node (16–64 GPUs) — Uses optical modules to reduce latency and improve cross-node communication efficiency for distributed training.

Yaolong S8000 G2 Super Node (32/64 GPUs) — Industry-first 3D Mesh interconnect enabling 64 C550 GPUs in a single super node; claimed 4× communication performance uplift; supports DeepSeek, Qwen, Kimi-K2, Step-3 and other mainstream large models.

Shanghai Cube High-Density Liquid-Cooled Cabinet (128 GPUs) — 47U cabinet holding four super nodes (32 GPUs each), eight cabinets form a 1,000-GPU cluster; full liquid cooling eliminates traditional air cooling, achieving industry-leading compute density and energy efficiency.

High-Density Liquid-Cooled Compute POD — Innovative architecture deeply integrated with MetaX clusters, supporting AI training, inference, and general compute with flexible scaling.

Core Super-Node Technologies

Proprietary Interconnect: MetaXLink super-node extension and 3D Mesh three-dimensional grid topology achieve low-latency all-to-all connectivity; any two GPUs communicate within two hops.

Unified Addressing & Memory Pooling: Remote GPU virtual mapping breaks multi-host switch-domain addressing limits, expanding GPU memory address space into a global unified memory pool to serve ultra-large model KV-cache demands.

Liquid Cooling & High-Density Design: Shanghai Cube and Compute POD adopt full liquid cooling, removing air cooling; single cabinet packs 128 GPUs, dramatically increasing compute density per unit space.

Chip & Ecosystem Compatibility: Powered by C550/C600 series chips supporting mixed-precision compute; compatible with PyTorch, TensorFlow and other mainstream frameworks; adapts to domestic GPU ecosystem and OAM/OCM modular upgrades.

Xijing S600 Super Node Deep Dive

The Xijing S600 is a system-level compute product targeting large-model training, inference, and intelligent computing center construction. It enables 64-GPU high-density deployment per cabinet via an all-to-all intra-cabinet interconnect fabric that drastically reduces data-exchange latency and boosts parallel computing efficiency for large models. The architecture flexibly supports expert parallelism (EP), tensor parallelism (TP), and other multi-dimensional parallel strategies, covering both training and inference scenarios.

Xijing S600 employs an OEX (Orthogonal Backplane-Free Interconnect) architecture and an innovative "Three-Zero Design" :

Compute Node Zero Cables: Internal node signal transmission requires no external cabling.

Switch Node Zero Cables: Switch layer uses orthogonal backplane direct connection.

Compute-to-Switch Zero Cables: Critical intra-cabinet data paths have zero external wiring.

This approach greatly simplifies traditional cabinet cabling complexity, reducing signal loss and failure risk.

On scaling, Xijing S600 supports both intra-cabinet high-speed interconnect and inter-cabinet horizontal expansion, allowing cluster sizes to scale flexibly to 10,000 GPUs, meeting large intelligent computing center requirements. The integrated whole-cabinet delivery model shortens deployment cycles, while the zero-cable direct-connect design between compute and switch nodes further lowers construction and O&M costs.