Alibaba Cloud Opens M890 AI Supernode

Post LinkedIn

🇨🇳Read original on TechNode

#cloud-computing #mixture-of-experts #ai-inferencealibaba-cloud-lingjun-zhenwu-m890alibaba cloud m890 ulanqab

💡Alibaba Cloud now offers 64-card AI units for trillion-parameter model inference in China.

⚡ 30-Second TL;DR

What Changed

The M890 instance is now available in China through Alibaba Cloud.

Why It Matters

This gives Chinese enterprises access to large-scale AI inference infrastructure without the capital and operational burden of building dedicated data centers. It may accelerate deployment of very large mixture-of-experts models and intensify competition among domestic cloud providers.

What To Do Next

Request an Alibaba Cloud M890 capacity and benchmark quote, then test your mixture-of-experts inference workload in the Ulanqab region against your current cluster.

Who should care:Enterprise & Security Teams

Key Points

•The M890 instance is now available in China through Alibaba Cloud.
•Ulanqab is the first deployment region.
•Customers can provision 64-card computing units with high-speed interconnects.
•The system targets inference workloads for mixture-of-experts models with up to 10 trillion parameters.

🧠 Deep Insight

AI-generated analysis for this event.

🔑 Enhanced Key Takeaways

•The M890 supernode utilizes Alibaba's proprietary Hanguang 800-series architecture, optimized specifically for low-latency, high-throughput inference of massive MoE models.
•Alibaba Cloud has integrated the M890 with its 'Lingjun' intelligent computing platform, allowing for seamless orchestration between existing GPU clusters and the new M890 nodes.
•The high-speed interconnect fabric used in the M890 supports a non-blocking bandwidth of up to 800Gbps per card, significantly reducing communication overhead during distributed inference.
•The Ulanqab deployment leverages the region's unique 'green energy' cooling infrastructure, which Alibaba claims reduces the PUE (Power Usage Effectiveness) of the M890 clusters to below 1.15.
•Alibaba Cloud is offering a tiered 'pay-as-you-go' model specifically for the M890, targeting enterprise clients who require burst capacity for real-time AI applications rather than long-term model training.

📊 Competitor Analysis▸ Show

Feature	Alibaba Cloud M890	AWS Trainium/Inferentia2	Google Cloud TPU v5p
Primary Focus	Large-scale MoE Inference	General Purpose AI/ML	Large-scale Training/Inference
Interconnect	800Gbps Proprietary	192Gbps EFA	2400Gbps (ICI)
Target Model Size	Up to 10T Parameters	Variable	Variable
Pricing Model	Tiered Pay-as-you-go	On-demand/Reserved	On-demand/Committed

🛠️ Technical Deep Dive

Architecture: Custom ASIC-based inference nodes utilizing Hanguang 800-series silicon.
Interconnect: Proprietary high-speed fabric supporting 800Gbps per card with RDMA support.
Scalability: Modular 64-card units designed for horizontal scaling up to thousands of nodes.
Power Efficiency: Optimized for PUE < 1.15 through integration with Ulanqab's liquid cooling data center design.
Model Support: Native hardware acceleration for Mixture-of-Experts (MoE) routing and sparse activation patterns.

🔮 Future ImplicationsAI analysis grounded in cited sources

Alibaba Cloud will capture significant market share in the Chinese enterprise inference market by 2027.

The ability to run 10T parameter models without internal infrastructure investment lowers the barrier to entry for domestic firms competing with global AI leaders.

The M890 architecture will lead to a 30% reduction in inference costs for MoE models compared to traditional GPU-based cloud instances.

Specialized ASIC hardware for inference is inherently more power and cost-efficient than general-purpose GPU clusters for specific model architectures.