Global GPU Computing News【20250830】

MIIT Emphasizes Orderly Guidance for Computing Facility Construction: At the 2025 China Computing Conference, the Ministry of Industry and Information Technology (MIIT) stated it would promote improvements to the policy system for computing layout, optimize computing infrastructure, guide regions in rational planning of intelligent computing facilities, and accelerate breakthroughs in key core technologies like GPUs.

Domestic AI Chips Collaborate on UE8M0 FP8 Standard: During the release of Deepseek 3.1, Deepseek announced that the next generation of domestic AI chips (e.g., Huawei Ascend, Moore Threads) will support the "UE8M0 FP8" algorithm standard. This standard aims to improve performance and reduce memory pressure and power consumption, creating a differentiated competitive path against NVIDIA's dominant NVFP4 approach.

Seven Computing Centers Join National Supercomputing Internet: At the 2025 China Computing Conference, seven computing centers, including the National Supercomputing Center in Taiyuan and the Qingdao "Haizhixin" AI Computing Center, officially joined the National Supercomputing Internet. This brings the number of connected backbone node-type computing centers to over 30, with coalition members exceeding 200.

NVIDIA Blackwell Ultra Details Revealed: NVIDIA disclosed that the upgraded version of its Blackwell GPU, "Blackwell Ultra" (B300 & GB300), will be released within the year. Key features include use of TSMC's 4NP process, a transistor count of 208 billion, support for PCIe 6.0 (doubling bandwidth), HBM3E memory increased to 288GB, but a maximum power consumption of 1400W.

GB300 Shows Significant Inference Efficiency Gains: Tests by cloud provider CoreWeave showed that the Blackwell-based GB200 chip (often referring to the superchip containing B200 GPUs, sometimes called GB200/GB300 in context) running the DeepSeek R1 inference model required only 4 GPUs to complete a task that originally needed 16 H100s, achieving a roughly 6x improvement in inference throughput.

NVFP4 Precision and Efficiency Stand Out: NVIDIA's NVFP4 data format, heavily promoted for Blackwell, showed minimal precision loss compared to FP8 (less than 1 percentage point behind in most scenarios) in tests, while reducing memory usage by 1.8x and improving energy efficiency by up to 50x compared to the H100 architecture.

NVIDIA Launches Spectrum-XGS Ethernet Platform: At the Hot Chip conference, NVIDIA introduced Spectrum-XGS, designed to efficiently interconnect multiple data centers across different geographical locations, enabling millions of GPU nodes to collaborate like a single "superchip", overcoming the physical and power limitations of a single data center.

India's First GPU Cloud Platform Deploys Blackwell Servers: Indian company Novacore Innovations announced the deployment of the country's first GPU cloud platform based on NVIDIA Blackwell servers (B200). Having secured $5.1 million in funding, it aims to provide lower-cost AI computing power to startups and enterprises in India, the US, and the Middle East.

NVIDIA Introduces Jetson Thor Robot "Brain": NVIDIA released the new-generation robot computing module Jetson Thor, featuring a Blackwell GPU. It delivers up to 2070 TFLOPS of peak AI performance at FP4 precision. Its AI computing performance is increased by up to 7.5x compared to the previous generation, with energy efficiency improved by up to 3.5x, aiming to advance humanoid robots and physical AI.

Next-Gen Rubin GPU Confirmed in Development: Alongside the details on Blackwell Ultra, NVIDIA also confirmed that its next-generation GPU architecture, "Rubin", is under development, signaling the continued rapid iteration of its GPU technology roadmap.