1,500 Tons and 800 Gb/s: China’s Sugon 8000 AI Supercluster in Detail

TVS News AI, editorially supervised, presents 1,500 Tons and 800 Gb/s: China’s Sugon 8000 AI Supercluster in Detail.

1,500 tons, 800 gigabytes per second: The more imposing the new data center in Zhengzhou sounds, the greater the challenge the Sugon 8000 has to overcome. The system is designed to bundle 100,000 accelerator cards, all of which were developed and manufactured in China. With it, the manufacturer aims not merely to operate an individual data center, but to build a central hub for the country’s national supercomputing infrastructure.

Xataka had already reported on the project, highlighting the cluster’s dimensions and high utilization. New are the details on networking, cooling and capacity shown during the national supercomputing conference in Shenzhen. The scale becomes more tangible when the cluster is understood not as an abstract supercomputer, but as a collaborative system comprising thousands of nodes, tens of thousands of cards and a cooling circuit built specifically for it.

A National Computing Hub With More Than 13,000 Nodes

The central hub

The Sugon 8000 was not built solely for the Chinese state. It forms the central hub of the National Supercomputing Internet, a nationwide network that allows universities, research institutes and companies to access computing power remotely. The project therefore concentrates major capacity at a single location and then makes it available to other institutions.

According to Sugon, the cluster operates with more than 13,000 nodes and is nearly fully utilized. At the same time, more than 6,000 tasks are being processed simultaneously. These figures come from Sugon and the Chinese state media cited by Xataka.

Chinese components at every level

The ambition extends far beyond graphics cards. According to the company, the chips, networking equipment, storage, cooling system and programming software all come from Chinese suppliers. This is particularly significant given the export restrictions on US technology: According to the US regulatory database, Sugon was added to the Entity List in 2019. As a result, exports, re-exports and technology transfers by US companies require prior authorization from the US Treasury Department.

The Real Problem Is Heat

100,000 chips at high temperatures

When all the chips operate at high speed, an enormous amount of heat is generated. According to the outlet that visited Sugon in Zhengzhou, the heat produced per hour is equivalent to roughly burning 12 tons of coal. Excessively high temperatures force the chips to reduce their performance. At the same time, the failure rate rises.

The technical answer is liquid cooling. The cards are immersed in a nonconductive fluorinated fluid that evaporates at approximately 50 °C. An external water system removes the resulting heat. The water circulating in the loop therefore does not replace the actual refrigerant circuit, but handles an additional part of heat dissipation.

Technology instead of ever more cooling surfaces

Especially with a cluster of this size, cooling is an integral part of the computing architecture. The advantage of this arrangement is that it directly cools the components. The downside is that the system requires an enormous, precisely coordinated infrastructure of pipes, sensors and fluid circuits. This also explains why Sugon treats the water as an external part of the system rather than an incidental element.

At 800 Gb/s, 100,000 Cards Are Supposed to Work Like One Machine

The network as a collective

Each network port transmits data at 800 gigabits per second. This high bandwidth is intended to make data exchange between the chips fast enough for the cluster to operate as a single shared computing machine. With 100,000 cards, this is decisive: An individual task must be able to distribute data across many nodes and then bring their results back together.

According to the engineers on site, the Sugon 8000 occupies an area of approximately 2,000 square meters. The system weighs around 1,500 tons and contains approximately 1,600 kilometers of internal cabling. These figures illustrate how the boundaries between the computer, the network and the building blur: The cable network, fluid pipes and spatial infrastructure all belong to the same supercomputer.

What China Already Plans to Do With the Cluster

More than 300 applications

Sugon says it has already optimized more than 300 applications in more than 30 fields of research and industry. The cluster is therefore intended not only to provide a large pool of computing resources, but also to accelerate scientific and industrial tasks.

The most tangible addition compared with the initial reports on the Sugon 8000 is the technical detail: The dimensions of 2,000 square meters and 1,500 tons are now joined by 1,600 kilometers of cable, a liquid cooling system at around 50 °C and 800 Gb/s per network port. Together, they portray a national data center whose performance depends not only on its chips, but above all on whether the networking, storage and cooling can handle the enormous flow of data.