Global GPU Computing News【20250426】

1. U.S. Restricts NVIDIA H20 Chip Exports, Accelerating Domestic Substitution
   The U.S. Department of Commerce added NVIDIA’s H20 chip to its export control list, leading to a $5.5 billion write-down for NVIDIA. The H20, a mainstream AI chip in China, had been heavily stockpiled by Tencent, ByteDance, and others. This restriction is accelerating adoption of domestic alternatives like Huawei’s Ascend and Hygon’s chips.

2. Huawei’s CloudMatrix 384 Super Node Cluster Deployed Commercially
   Huawei launched its Ascend-based CloudMatrix 384 super node cluster, delivering 300 PFlops of computing power—67% higher performance than NVIDIA’s NVL72 at 53% lower cost. The cluster is now operational in Wuhu’s data center, supporting large models like DeepSeek-R1 with inference performance rivaling the H100.

3. Beijing Releases Computing Infrastructure Plan, Prioritizes Domestic GPUs
   Beijing’s Computing Infrastructure Construction Plan (2024–2027) aims to build large-scale AI clusters, offering subsidies for companies purchasing domestic GPUs. The target is to reach 45 EFLOPS of AI computing capacity by 2025.

4. Moore Threads’ "Kuae" 1,000-Card Cluster Deployed
   Moore Threads completed deployment of its 1,000-card AI clusters in Beijing (Yizhuang, Miyun) and Nanjing. These clusters support training for 100-billion-parameter models with 91% scaling efficiency, CUDA compatibility, and sub-second fault recovery.

5. Domestic Ecosystem Gains Momentum
   Huawei Ascend, Hygon, and others are narrowing the gap with NVIDIA through system-level innovations (e.g., chip clusters, high-speed interconnects). Domestic AI chips are projected to grow from 12% to 34% market share by 2026.

6. NVIDIA H20 Supply Chain Risks Escalate
   Major Chinese clients like Tencent and ByteDance face dwindling H20 inventories. H3C warns of production and logistics disruptions, driving further demand for domestic alternatives.

7. Global LLM Training Fuels Demand for 10,000+ GPU Clusters
   Meta trained Llama 3 on a 24,000-GPU cluster, while Elon Musk’s Grok 3 will reportedly require 100,000 H100 GPUs. Multi-thousand-card clusters are now standard for LLM training, with interconnect efficiency becoming critical.

8. Breakthroughs in Heterogeneous Computing & System Integration
   Innovations like chiplet integration, liquid cooling, and advanced interconnects are driving performance gains more than single-chip improvements.

9. LanYun Launches "OriginAI" Intelligent Computing Platform
   The platform integrates GPU resource scheduling and AI development tools, enabling rapid model deployment. It also offers "1 million computing vouchers" to incubate 100 AI projects.

10. Huawei Ascend Supports 160+ Major LLMs
    Huawei’s Ascend Cloud services now power 400+ applications in finance, healthcare, and other sectors. For example, a top-tier hospital reduced CT scan analysis time from 30 minutes to 5 seconds.

Summary

This week’s highlights include China’s computing breakthroughs (Huawei, Moore Threads), policy support (Beijing’s plan), and global supply chain shifts (H20 restrictions). Domestic GPUs are leapfrogging through cluster technology and ecosystem adaptation, while global LLM demands push cluster scalability.