🧠 Food for thought

Implications, context, and why it matters.

The scale of China’s new AI cluster could outpace today’s top supercomputers

• The Zhengzhou system combines three of Sugon’s scaleX platforms. It began trial runs on Thursday, SCNet said, along with Sugon 1. SCNet is a national computing network that links computing resources for researchers and companies.
• Sugon calls scaleX China’s first 10,000-card AI supercluster. The company puts total computing power at more than 5 exaflops 2.
• Three platforms put potential performance near 15 exaflops, using a single scaleX as the baseline. That would beat El Capitan’s maximum performance by more than eight times on the latest TOP500 list of the world’s fastest supercomputers 1.
• Each platform uses scaleX640 supernodes. Sugon describes them as the world’s first single-cabinet 640-card supernode 2.

Sugon’s open architecture signals a push for a self-reliant AI ecosystem

• Sugon’s hardware connects to China’s state-backed SuperComputing Network (SCNet) 1.
• At the hardware layer, the scaleX640 supernode supports accelerator cards from multiple brands. Cailian Press says it also works with mainstream computing ecosystems at the software layer 3.
• That setup puts Sugon up against Chinese rivals like Huawei and Alibaba, which have built their own supernode products for the same market 3.
• SCNet aims to pair hardware from companies such as Sugon with domestic software platforms. The effort targets a more self-reliant AI development ecosystem from chips to software 1.