Global GPU Computing News【20251122】

Huawei Releases Flex:ai Container Software: On November 21, Huawei released and open-sourced its Flex:ai AI container software. This technology can partition a single GPU/NPU computing card into multiple virtual units as small as 10%, enabling a single card to simultaneously handle multiple AI tasks. It is claimed to increase typical GPU/NPU utilization from 30%-40% to 70%.

Changsha High-Performance Computing Research Center Established: On November 21, the Changsha High-Performance Computing Research Center was signed and settled in the Jinxia Economic Development Zone during the 2025 World Computing Conference. The center will put into production the high-performance GPGPU chip "Quanyingxing Q3", with performance targets aiming to catch up with and surpass NVIDIA's A100, primarily targeting scientific computing and cloud computing.

Moore Threads Initiates STAR Market Subscription: Domestic GPU company Moore Threads will start its subscription on the Sci-Tech Innovation Board (STAR Market) on November 24, expecting to raise nearly 8 billion yuan. The company acknowledged gaps in product performance and ecosystem development compared to international leaders, but stated that the single-precision floating-point computing power of some of its products is close to that of the NVIDIA RTX 3060.

NVIDIA to Deploy New Supercomputers for Japan's RIKEN: NVIDIA is collaborating with Japan's RIKEN institute to deploy two new supercomputers based on the Blackwell architecture, utilizing a total of 2,140 Blackwell GPUs. One will be used for AI research in life sciences, materials science, and other fields, while the other will be dedicated to quantum computing. They are expected to be operational in Spring 2026.

NVIDIA Introduces Apollo Physics Simulation AI Model: At the SC25 supercomputing conference, NVIDIA launched a product called Apollo, an AI model for physical simulation. It can compress years of simulation data and knowledge into AI models that run in real-time, aiming to provide optimization for specific physical workloads.

HMCI and Rapt.AI to Deploy NVIDIA GB10 Systems: HMCI and Rapt.AI announced the deployment of NVIDIA GB10 systems to build an AI and robotics ecosystem for academia, startups, and municipal teams. The project, funded by a $5 million investment from the City of Rancho Cordova, USA, is expected to launch in early 2026, offering both remote and on-site access.

France to Build its First Exascale Supercomputer: French company Eviden, in collaboration with AMD, will build France's first exascale supercomputer, Alice Recoque. The system will use AMD EPYC CPUs and Instinct MI430X GPUs, delivering performance exceeding 1 exaflop, and emphasizes approximately 20% lower power consumption compared to similar systems.

Dell Releases Multiple AI Servers and Switches: At the SC25 conference, Dell Technologies launched several servers designed for AI training and inference, including the XE9785 series using AMD Instinct accelerators and the PowerEdge R770AP server based on Intel. It also introduced two new switches providing 102.4 Tb/s switching capacity to handle AI workloads.

Sustainable AI Infrastructure to be Located in Nordic Region: Sustainable data center provider Verne and hyperscale company Nscale announced plans to deploy a 15-megawatt AI infrastructure in Iceland. The facility is scheduled to be equipped with approximately 4,600 NVIDIA Blackwell Ultra GPUs in 2026, with 85% utilizing liquid cooling and leveraging 100% local renewable energy.

NVIDIA Reports Strong Financial Results: On November 20, NVIDIA reported third-quarter revenue of $57 billion, a 62% year-over-year increase. Data center segment sales reached $51 billion, up 66% year-over-year. The company's CFO stated that the visible pipeline of related revenue for its Blackwell and Rubin platforms from the beginning of this year through the end of 2026 amounts to $500 billion.