Alibaba Cloud Redefines Cloud AI with 64-GPU Super Node Instance

At the World Artificial Intelligence Conference, Alibaba Cloud made a significant leap in cloud infrastructure by introducing the Lingjun Zhenwu M890 Super Node instance. This offering represents a first for public cloud: a fully integrated, high-performance computing unit comprising 64 GPUs, delivered as a ready-to-use service.

Architectural Leap: Unlocking Cluster-Wide Performance

The true bottleneck in large-scale AI isn't just raw compute power, but how efficiently hundreds of chips can communicate. The M890 Super Node tackles this head-on with a groundbreaking interconnect design.

Powered by the in-house developed ICNSwitch1.0 chip, the instance scales up GPU interconnectivity from 16 to a cohesive cluster of 64. The inter-GPU bandwidth reaches a remarkable 800GB/s. This high-bandwidth, low-latency fabric minimizes communication overhead, ensuring data flows seamlessly across the cluster during intensive model operations and maximizing usable computational output.

Engineered for Trillion-Parameter Model Era

As AI models evolve into trillion-parameter behemoths, particularly with Mixture-of-Experts architectures, infrastructure demands grow exponentially. The Zhenwu M890 is built for this next generation.

It features native support for FP8 and FP4 low-precision compute formats. This capability allows for significant reductions in memory footprint and power consumption while maintaining model accuracy, leading to superior efficiency. Alibaba Cloud states that a single M890 Super Node instance is capable of handling inference for MoE models scaling to trillions of parameters, providing a robust cloud foundation for deploying state-of-the-art AI.

Turnkey Simplicity for Complex AI Workloads

Managing a large GPU cluster traditionally involves immense complexity in hardware, networking, and software. The "Super Node" concept delivers a simplified, consumption-based model.

  • Integrated Solution: Users get a pre-configured, high-performance compute block, eliminating the need for complex cluster assembly and tuning.
  • Cloud-Native Integration: As a public cloud service, it seamlessly connects with Alibaba Cloud's storage, database, and networking ecosystems.
  • Operational Agility：Organizations can access this cutting-edge power on-demand, accelerating innovation cycles without the capital expenditure and lead times of building private infrastructure.

The Lingjun Zhenwu M890 Super Node instance is now available for invitation-based testing in the Ulanqab region. This move not only provides new options for tech giants but also democratizes access to elite AI computing, enabling a broader range of enterprises and researchers to bring ambitious AI projects to life.