Product Highlights

• 

  Extreme Performance

  Kilocard-scale compute units with extreme AI compute

  A 1-to-2 high-density design interconnects 640 accelerators in one cabinet over an ultra-high-speed bus, forming a large-scale, high-bandwidth, low-latency super-node domain. Two ScaleX640 super nodes make a 1,280-card compute unit, linked between cabinets over a high-speed network. The CDM liquid heat-exchange unit delivers up to 1.72 MW of cooling for kilocard-scale compute.

• 

  Peak Efficiency

  Pushing model, energy, and space efficiency to the limit

  Hardware-software co-optimization raises MoE training efficiency and high-throughput inference by 30–40% versus traditional architectures. Ultra-high-speed orthogonal design, high-density blades, immersion two-phase liquid cooling, and high-voltage DC power combine into one of the industry’s most integrated liquid-cooled super nodes. Cabinet compute density can be up to 20× higher than other super nodes, with PUE as low as 1.04.

• 

  Fully Open

  Flexible hardware choices and seamless software migration

  An open AI compute architecture supports multi-vendor AI accelerators. It is compatible with mainstream AI software stacks, optimized for 400+ foundation models, and lets AI models and business applications migrate and be tuned with minimal friction.

• 

  Ultra Reliability

  Multi-level reliability for 100,000-card clusters

  From single-node RAS to cluster operations and fault recovery, multi-layer reliability has been proven in 30+ days of stability testing to support 100,000-card-scale deployment.