Alibaba Panjiu AL64 Super Node: 500K-GPU Scale Architecture with Zhenwu V900 Chip

Panjiu Super Node Product Roadmap (2026–2028)

The 2026 Yunqi Conference revealed a four-stage roadmap for Panjiu super nodes:

2026 – AL128 (Gen 1): 128 AI accelerator cards per cabinet, 350 kW power density requiring liquid cooling, 150 ns ultra-low latency, petabyte-scale Scale-Up bandwidth, supports Zhenwu M890.

2027 – AL128 (Upgrade): Retains 128-card form factor, adds 800 V HVDC (high-voltage DC power) for improved supply efficiency, supports Zhenwu V900 and other GPUs.

2027 – AL64 (Distributed Architecture): 64 cards per cabinet, distributed cabinet deployment (unlike prior high-density single cabinets), Scale-Up to 1,024-card clusters; adopts SNPO all-optical interconnect to reduce long-distance communication latency, supports Zhenwu V900.

2028 – AL144 (Ultimate Form): 144 cards per cabinet, 650 kW (double power), Scale-Up to 10,368 cards (massive cluster expansion), doubled Scale-Up bandwidth, vertical micro-channel + liquid metal cooling (advanced thermal), native 800 V HVDC, supports Zhenwu J900 and other GPUs.

Panjiu AI-Native New-Generation Super Node Architecture

The Panjiu AL64 Zhenwu V900 super node, paired with Panjiu AI Infra 3.0, targets Agent-era large-model training and ultra-large-cluster inference. It achieves full-stack integration of self-developed compute, interconnect, NIC, and storage controller chips, natively supports wide-area super nodes, and scales a single AI cluster to 500,000 cards. Key architectural features:

Orthogonal Decoupled Architecture: CPU head and XPU tail separated; three blind-mate modular design enables compute, switch, and storage module replacement without cabinet removal, cutting maintenance time from hours to minutes.

Single-Node Capacity: 64 Zhenwu V900 cards; Scale-Up non-convergent expansion to 1,024 cards within a single super-node cabinet domain.

Three-Layer Interconnect Fabric

Scale-Up Domain (Intra-Cabinet): ICN Switch 2.0, thousand-card full bandwidth, native memory semantics, unified addressing.

Scale-Out Domain (Inter-Cabinet): Panmai 920 smart NIC + high-speed optical network for horizontal super-node expansion.

Wide-Area DCN Domain (Multi-DC Cross-Region): Alibaba Cloud next-gen intelligent computing network enabling wide-area super nodes, cluster ceiling of 500,000 cards.

Thermal & Power

Full liquid cooling; single compute module up to 3,500 W; cabinet supports high-power high-density deployment.

Hardware Interfaces

SNPO near-package optical interconnect, 51.2 Tbps optical switching, 64 × 800G optical ports; copper/optical shared socket, compatible with future 224G/448G evolution.

Deep Integration of Four Self-Developed Core Chips

The new Panjiu super node fuses T-Head's full-stack custom silicon:

Zhenwu V900 AI Accelerator: Next-gen train/inference unified chip; 3× comprehensive performance vs. previous M890 flagship; 216 GB VRAM; 1,200 GB/s inter-chip bandwidth; native FP8/FP4 low-precision compute; targets trillion-to-10-trillion parameter model training and inference; slated for Q1 2027 volume production.

ICN Switch Interconnect Chip: Custom high-speed interconnect switch; handles thousand-card full-bandwidth intra-super-node communication; reduces cross-card latency.

Panmai Smart NIC: Offloads Scale-Out network protocol processing from CPU to NIC, drastically lowering host overhead and boosting multi-node communication efficiency for large-scale distributed training.

Zhenyue SSD Controller: Custom storage controller optimized for AI workloads — KV-Cache, model weights, dataset high-speed read/write; bridges compute and storage paths, alleviating large-model inference storage bottlenecks.

Compared with the previous Panjiu AL128 (M890) which already validated 2-trillion-parameter model training, the new Panjiu + Zhenwu V900 pushes the model-scale ceiling higher while improving cluster compute utilization.

T-Head Zhenwu V900 AI Chip Details

The conference underscored a shift from "single-point compute" to "full-stack system" competition. T-Head committed to a "one generation per year" cadence: 2024 Zhenwu 810E (96 GB VRAM), 2026 M890 (144 GB VRAM), 2027 Q1 Zhenwu V900 (216 GB VRAM). Zhenwu V900 is built on a custom parallel compute architecture, delivers 216 GB VRAM, 1,200 GB/s inter-chip bandwidth, native FP8/FP4 support, raising compute density while significantly cutting inference cost.

Wide-area super nodes are not the endpoint. Post-V900, Zhenwu J900 will further break compute ceilings, and Alibaba Cloud's intelligent computing network will continue evolving to enhance cross-region wide-area compute collaboration.