Global GPU Computing News【20250913】

NVIDIA Launches Dedicated Inference GPU Rubin CPX: NVIDIA introduced the Rubin CPX GPU, specifically designed for ultra-long-context inference. It features a monolithic design, equipped with 128GB of GDDR7 memory, and delivers 30 PetaFLOPS of computing power at NVFP4 precision. It is focused on tasks requiring processing of millions of tokens, such as codebase analysis and long video generation. Its attention mechanism performance is claimed to be 3 times that of the current flagship GB300 NVL72 system.

Rubin CPX Employs Decoupled Inference Architecture: This chip aims to separate the context processing (compute-intensive) and content generation (memory bandwidth-intensive) stages of AI inference. Through targeted optimization, it is expected to improve overall inference efficiency by up to 6.5 times.

Building the Vera Rubin Platform: The Rubin CPX will form the NVIDIA Vera Rubin NVL144 CPX platform together with the Vera CPU and Rubin GPU. A single rack is expected to deliver 8 Exaflops of AI compute (NVFP4), 100TB of high-speed memory, and 1.7 PB/s memory bandwidth. Its AI performance is claimed to be 7.5 times that of the GB200 NVL72.

Rubin CPX Expected Launch Date: The Rubin CPX is expected to be available in late 2026.

China's First Open AI Computing Architecture Released: Sugon (Dawning Information Industry), in collaboration with over 20 industry chain companies, released China's first Open AI Computing Architecture and the Sugon AI Super-Cluster System at the 2025 World Intelligence Industry Expo. It aims to address computing bottlenecks through open collaboration, building a collaborative innovation system covering "computing, storage, networking, power, cooling, management, and software."

Performance Boost in Sugon AI Super-Cluster System: The system supports 96 accelerator cards per cabinet, providing hundreds of petaflops of computing power, scalable to a million-card scale. Officially, its thousand-card cluster training and inference performance reaches 2.3 times the industry mainstream level, model development efficiency improves by 4 times, GPU computing efficiency increases by 55%, and fault repair time is shortened by 47%.

Emphasis on Openness and Compatibility: This architecture strives to be compatible with multi-brand GPU accelerator cards and supports mainstream software ecosystems like CUDA, aiming to reduce user software development and adaptation costs and avoid lock-in to a single technology路线 (technology path/roadmap).

Opening Key Technical Capabilities: Sugon announced the opening of three core technical capabilities: AI storage optimization technology, liquid cooling infrastructure design specifications (its market share in immersion liquid cooling exceeds 50%), and the DeepAI basic software stack, to promote ecosystem co-construction.

Establishment of a Joint Laboratory: Relying on the National Advanced Computing Industry Innovation Center, a "Joint Laboratory for Open AI Computing Architecture" was established in partnership with chip, OEM, and other companies. It will focus on five types of technology openness at the component and system levels, aiming to reduce redundant R&D.

Lisuan TrueGPU Cards Enter Mass Production: Lisuan Technology announced that its first fully self-developed high-performance graphics GPU, the "7G100" series, and the Lisuan eXtreme series graphics cards, will start mass production in September. Official demonstrations showed an average frame rate of over 70 fps in "Black Myth: Wukong" at 1080p high settings, and it is claimed to meet the needs of personal AI PCs for certain AI models.