Alibaba Cloud Super Node: 350kW Racks, 224G Interconnect, Biomimetic Liquid Cooling

Challenges in Scale-Up High-Density Interconnect

Large model training demands ultra-low latency, high-reliability Scale-Up interconnect. At 112Gbps PAM4 and 224Gbps, signal integrity (SI) pressure increases dramatically across packaging, PCB traces, backplanes, and high-speed cables.

1. High-Speed Interconnect: 112G/224G Signal Integrity

Signal traverses multiple physical media, each adding loss.

2. Power Delivery: Chip and Rack Power Leap

GPU module power escalated from 100W to 4900W; single AI chip exceeds 2000W. Rack power reaches 350kW+ . Traditional distributed power architecture cannot meet capacity, reliability, and maintainability demands.

3. Cooling: Liquid Cooling Reliability for 2000W+ Chips

Air cooling hits ceiling; cold-plate liquid cooling becomes mandatory. Challenges: cold-plate welding seal reliability, leak protection, explosion/splash prevention, fault warning. A chip-node-rack three-level leak detection and response system is essential.

Panjiu AI Infra 2.0 Super Node Design

Alibaba Cloud's Panjiu AI Infra 2.0 super node (AL128) targets high density, high reliability, open architecture.

High-Density Interconnect: Single rack supports up to 128 GPUs, enabling 64-GPU Scale-Up high-speed interconnect.

High-Reliability Operation: High-speed orthogonal interconnect + modular design for minute-level component replacement; system power achieves 6 nines (99.9999%) reliability; three-level leak detection at chip, node, and rack levels.

Open Architecture: Self-developed ALink interconnect protocol decouples GPU, CPU, and ALink switch nodes; standardized hardware interfaces support multi-vendor GPUs, CPUs, and switch chips, avoiding single-vendor lock-in.

Core challenges: 112G/224G signal integrity with low-cost copper; 2000W chip power delivery and 350kW+ rack power; 2000W+ chip cooling and cold-plate system reliability.

Key Engineering Innovations

1. Ultra-High-Power Centralized Power Architecture

Rack-level centralized power delivers 350kW+ with high reliability.

ATS Dual-Input Power: Optimized main load range, PSU efficiency >97%; built-in large capacitors handle EDPP power-loss protection scenarios.

Passive PowerShelf: Passive hardware design, 10-year design life at 55°C; supports front-access blind-mate hot-swap for online maintenance without downtime.

Zoned BusBar Copper Bars: Pooled zoned power distribution, optimized current density; digital voltage-drop detection monitors BusBar connection status in real time.

Intelligent SMC Management: Custom management chip enables multi-PowerShelf high-precision current sharing; supports online hot-swap, remote monitoring, batch operations.

2. Biomimetic Leaf-Vein Adaptive Flow Liquid Cooling

Addresses 2kW per chip and 350kW+ per rack heat dissipation.

Biomimetic Multi-Level Branching Pipes: Adaptively allocates coolant flow to match per-chip heat variation, maximizing cooling capacity.

Fully Welded Cold Plates + Explosion/Splash Protection: Zero-leak cold plates extend lifespan; rack-level explosion and splash protection mitigates secondary risks from pipe failures.

Three-Level Leak Detection (Chip, Node, Rack): Hierarchical real-time monitoring, graded fault response, predictive failure alerts for predictable maintenance.

3. GPU and ALink Switch Orthogonal Interconnect

Balances high-density high-speed interconnect with maintainability.

Fully modular interconnect; faulty parts replaced in minutes.

224G all-copper orthogonal interconnect reduces link loss.

Partitioned independent sizing supports orthogonal multi-bus blind-mate mating.

Future Evolution Directions

Interconnect: Advance 224G SerDes , explore next-gen copper performance limits; evolve orthogonal architecture toward higher node density.

Power: Move toward megawatt-level rack power; explore 800V DC high-voltage DC rack power architecture.

Cooling: Fuse chip-level cooling with system-level liquid cooling; pursue 100% heat capture toward fan-less whole-system design.

Next-Generation Panjiu Super Node (AL64 / Zhenwu V900)

Integrates Zhenwu V900, ICN Switch, Panmai smart NIC, and Zhenyue SSD controller. Combined with Alibaba Cloud's new intelligent computing center network architecture, achieves compute-storage-network full-stack system collaboration, scaling single AI cluster to 500,000 GPUs , evolving toward wide-area super nodes.

The competition for next-gen super nodes is fundamentally a dual contest of engineering depth and standards breadth , requiring one-generation-ahead technology layout across interconnect, power, and cooling.