Larger models require more hardware to train. China responded with a supercluster of 100,000 national cards

A few months ago we had the chance to talk about the exploits of the Sugon 8000, a supercluster with around 100,000 graphics cards, all designed and manufactured in China. The system is located in Zhengzhou, Henan Province and was presented to the public during China’s National High Performance Computing Conference in Shenzhen, showing more details of its operation.

The pride of the nation. Building a machine of this size means chips, networking, storage, cooling and programming software to work together, and Sugon says every layer comes from Chinese suppliers. That’s important in a country whose tech industry has been under pressure from the United States for years. Specifically, according to the US Federal Register, Sugon was added to the Entity List in 2019, a list that restricts companies from exporting, re-exporting and transferring US technology without a prior license.

Details. The cluster is the central node of China’s National Supercomputing Internet, a nationwide network that provides remote access to computing power to universities, research institutes and enterprises. Sugon claims the cluster runs more than 13,000 nodes and is near full load, handling more than 6,000 jobs at a time, according to Chinese state media reports.

Site engineers explained to local media Hebccw that the system occupies about 2,000 square meters, weighs about 1,500 tons and contains about 1,600 kilometers of internal wiring.

"We are kept anonymous": Programmers have another front with AI: their mental health

heat problem Running 100,000 chips at full speed generates a lot of heat. In fact, according to the media, production per hour is comparable to burning 12 tons of coal. Excessive heat causes chips to slow down and increase failure rates, so Sugon immerses its boards in a non-conductive fluorinated liquid that boils at about 50 degrees Celsius. The heat is subsequently dissipated through an external water system.

Each network port transfers data at 800 Gb/s, so the chips can exchange information fast enough to operate as a single machine.

What you are already doing. Sugon claims to have optimized more than 300 applications in more than 30 research and industrial fields, from new materials and drug discovery to aerospace and agriculture. According to the South China Morning Post, examples differ in meteorology, where in September, the China Meteorological Administration and Sugon announced a 10-day global weather forecast with 5 km resolution, which the system produced in about an hour. It is more accurate than most major forecasting systems in general use.

Another notable example was that of Tsinghua University researchers, who used the cluster to reconstruct a digital model of the Yellow River’s sediment and water level behavior shared by interesting engineering.

Supercomputers for everything. The company presents the Sugon 8000 as a hybrid system, with traditional supercomputer behavior that prioritizes high-precision calculations for scientific simulations, and as an AI system, which sacrifices some accuracy for speed in processing large data sets. All on the same hardware.

Cover image good luck

In Xataka | OpenAI claims to have made progress on three very complex scientific problems, although academia will have the final say

Add a Comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Keep Up to Date with the Most Important News

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use
Wired Terra