Global GPU Computing News【20250809】

1. Moore Thread Launches Domestic GPU 7G100
   Performance rivals international mid-to-high-end GPUs (between RTX 4060 and 5060), achieving >70 FPS in Black Myth: Wukong at 1080p high settings. Features proprietary instruction sets and software stack, eliminating reliance on external IP.

2. Huawei Open-Sources CANN AI Architecture
   A CUDA alternative, CANN opens Ascend chip operator libraries and toolchains, supports mainstream frameworks like PyTorch, and lowers migration barriers to accelerate domestic AI ecosystem growth.

3. Huawei Unveils 384-NPU SuperNode Technology
   Enables full optical interconnect across 384 NPUs with 269 TB/s bandwidth, doubling NVIDIA GB200 system performance and breaking large-model training bottlenecks.

4. Huawei OceanStor A800 Storage Breaks World Record
   A single 8U device supports 255 H100 GPUs for training, delivers 698 GiB/s bandwidth, and ensures >90% GPU utilization—setting new standards for large-model checkpointing.

5. Greenland HK Launches 10k-GPU AI Center
   Partnering with Huajun Fund and Linxing Tech, "Shaolv AI Center" will deploy 10,000 GPUs within 3 years to serve the Greater Bay Area’s digital economy.

6. OpenAI Reveals GPT-5 Compute Infrastructure
   Trained on >200,000 GPUs; compute scaled 15x in 60 days to support 700M users. Next-gen 4.5-gigawatt data centers planned.

7. NVIDIA & OpenAI Release Open-Source Inference Models
   Launched gpt-oss-120b/20b with open weights, optimized for Blackwell’s 4-bit precision. Hits 1.5M tokens/sec inference speed, fully CUDA-compatible.

8. NVIDIA Upgrades NVLink Fusion
   Boosts cloud GPU scalability to 1M+ GPUs via NVLink Switch, enabling 1.8 TB/s bidirectional bandwidth.

9. High-Speed Interconnect Tech Diversifies
   Huawei’s optical push forces copper upgrades (224G DAC/AEC). NVIDIA adopts "copper-where-possible" strategy for short-range links.

10. Domestic Chips Gain Market Share
    Policy drives Huawei Ascend to 75% domestic AI chip output & 79% AI center share. NVIDIA’s China share drops to 50% amid localization surge.

Weekly Highlights

China advanced in indigenous architectures (e.g., Moore Thread), hyperscale clusters (Huawei 384-NPU/Greenland 10k-GPU), and open ecosystems (CANN). Globally, focus shifted to supercomputing efficiency (OpenAI) and inference optimization (NVIDIA). Optical vs. copper interconnect rivalry underscores foundational innovation in compute infrastructure.