摩尔线程GPU跑满了！无锡万卡级智算中心一期2000卡满载：20多家客户排队进场

Kuai Technology reported on November 2 that on October 31, Moore Thread signed a strategic cooperation agreement with Wuxi Huishan Investment Group and Neolithic Huitong. Relying on the Wuxi (Huishan) Domestic Intelligent Computing Center, Neolithic will carry out larger-scale autonomous driving model training based on full-featured Moore thread GPUs. The computing power demand is close to the 10,000-ka level, which is directly aligned with the 10,000-ka level expansion plan of the intelligent computing center. The first phase of the Wuxi (Huishan) Domestic Intelligent Computing Center with 2,000 cards was fully loaded in just three months after it was put into operation, setting a new record for the acceleration of the implementation of domestic commercial intelligent computing clusters. At present, more than 20 customers have conducted testing or formal business on the platform, of which training business accounts for more than 62%. A business structure that pays equal attention to reasoning and training has been formed. The core that carries the training load is the MTT S5000 integrated training and push computing card. The card is based on the fourth-generation MUSA "Pinghu" architecture. The AI ​​dense computing power of a single card can reach up to 1000 TFLOPS. It is equipped with 80GB of video memory, a video memory bandwidth of 1.6TB/s, and an interconnection bandwidth of 784GB/s between cards. It fully supports full-precision calculations from FP8 to FP64, and is compatible with mainstream AI ecosystems such as PyTorch and vLLM. Neolithic has completed the verification of autonomous driving model training based on this card, and the results have met expectations, so it decided to advance the training scale to the Wanka level. On the same day, the Neolithic East China Intelligent Unmanned Vehicle Headquarters project was launched in the Huishan Economic Development Zone, with a total planned investment of 1 billion yuan to build a four-in-one regional headquarters that integrates sales, settlement, operation scheduling, and intelligent manufacturing R&D. The goal of the three-party cooperation is not only to provide computing power, but to open up the entire chain from computing power supply, algorithm iteration, vehicle manufacturing to scene operation, so that domestic GPUs can move from "can run models" to "can support mass production." The first phase is fully loaded, training accounts for more than half, and the demand for 10,000 cards has been put on the table. This domestic intelligent computing center in Wuxi is verifying one thing: a full-featured GPU is not only a supplement to the inference end, but also has the ability to carry training-level tasks in complex loads such as autonomous driving that also involve AI computing, graphics rendering and physical simulation.