Skip to main content

Alibaba Cloud's 64-Card M890 Super Node Goes Live

Alibaba Cloud has officially rolled out its Lingjun Zhenwu M890 super node instance, and it's now up for grabs in the Wulanchabu region. What does that mean for businesses? Well, instead of pouring resources into building and maintaining their own data centers, enterprises can simply tap into this cloud-based computing behemoth. The M890 is designed to handle some seriously heavy lifting—think MoE (Mixture of Experts) large models with up to ten trillion parameters. That's not just a step up; it's a leap.

This isn't just another incremental upgrade. The M890 is the first super node in China to successfully run a large model with over two trillion parameters. In fact, Kimi K3 and Qwen3.8Max are already being served through this very instance. So, what's the secret sauce? For starters, it supports FP8 and FP4 low-precision computing, which speeds things up without sacrificing accuracy. But the real magic lies in the ICN Switch1.0 chip. This little marvel has expanded the Scale-up interconnection from a modest 16 cards to a whopping 64 cards, delivering an inter-card bandwidth of 800 GB/s. To put that in perspective, in training scenarios like intelligent driving and embodied intelligence, the M890 performs three times better than its predecessor, the Zhenwu 810E.

But the M890 doesn't exist in a vacuum. It's part of the broader Alibaba Cloud Lingjun Intelligent Computing Platform, which is built on the HPN8.0 training and inference integrated network. This platform can support up to 130,000 heterogeneous computing cards in a single cluster, and it's scalable to millions of cards. That kind of scale is what makes large-scale model training and inference not just possible, but practical.

The launch of the M890 marks a significant shift: super node computing power is moving from self-built infrastructure to the public cloud. As trillion-parameter and even ten-trillion-parameter MoE models start to roll out in earnest, the ability to interconnect cards at high speeds, leverage low-precision computing, and manage massive clusters becomes the new battleground in AI infrastructure. Alibaba Cloud is clearly staking its claim.

For businesses, this means the barriers to entry for cutting-edge AI are coming down. You no longer need to invest in expensive hardware or wait months to set up your own supercomputing facility. With the M890, you can get started almost immediately, scaling up as your needs grow. It's a democratization of sorts, and it's happening right now.

So, whether you're a startup dabbling in AI or an established enterprise looking to push the envelope, the M890 is worth a serious look. The future of AI is here, and it's more accessible than ever.

Key Points

  • Alibaba Cloud's M890 super node is now available in the Wulanchabu region, offering 64-card interconnect with 800GB/s bandwidth.
  • It's the first in China to run models with over two trillion parameters, supporting up to ten trillion parameter MoE models.
  • Performance boost: Three times faster than the previous generation in training scenarios like intelligent driving and embodied intelligence.
  • Part of a larger platform: The Lingjun Intelligent Computing Platform supports up to 130,000 cards per cluster, scalable to millions.
  • Shift to public cloud: Super node computing power is now accessible via cloud, reducing the need for self-built infrastructure.