Global GPU Computing News【20250726】

️ 1. Huawei Ascend 384 HyperNode Makes Global Debut

Huawei showcased its Ascend 384 HyperNode for the first time at WAIC 2025. Integrating 384 Ascend NPUs and 192 Kunpeng CPUs, it delivers 300 Pflops of computing power—1.7× faster than NVIDIA’s NVL72. Its network bandwidth (269 TB/s) and memory bandwidth (1229 TB/s) surged by 107% and 113%, overcoming bottlenecks in communication, cooling, and software adaptation.

⚡️ 2. Chinese GritTech G100 GPU Outperforms NVIDIA RTX 4060

GritTech’s G100 GPU appeared on Geekbench with an OpenCL score of 111,290, surpassing the RTX 4060 by ~10%. Built on the self-developed TrueGPU architecture, it supports DirectX 12 and other mainstream APIs. Official launch set for July 26.

️ 3. StepFun’s Step-3 LLM Optimized for Domestic Chips

StepFun released its Step-3 large model, achieving 300% higher inference efficiency on Chinese chips compared to DeepSeek-R1. It also co-founded the "Model-Chip Ecosystem Alliance" with Huawei Ascend, MetaX, and nearly 10 other chipmakers to advance domestic computing collaboration.

4. Moore Threads: LLM Compute Demand Soared 1000× in 5 Years

Wang Hua, VP of Moore Threads, revealed that large model training compute needs grew nearly 1000× between 2020–2025. FP8 precision and 10,000-GPU clusters are now essential. Moore Threads offers a domestic cluster solution supporting FP8 mixed-precision training, boosting performance by 20%–30%.

5. OpenAI’s Reasoning Model Wins IMO Gold, GPT-5 Teased

OpenAI’s new reasoning model reached gold-medal level at IMO 2025, solving 100-minute complex proofs. GPT-5 was previewed, highlighting breakthroughs in compute scaling and reinforcement learning.

6. NVIDIA Jetson Thor Edge Chips Enter Mass Production

NVIDIA’s edge-computing chips Jetson Thor T5000 (2070T FP4) and T4000 (1200T FP4) are set for shipment. Designed for embodied AI and real-time edge processing, they power autonomous driving and robotics with cloud-edge synergy.

7. NVIDIA/AMD Resume "Downgraded" GPU Sales to China

NVIDIA’s H20 and AMD’s MI309 "China-compliant" GPUs resumed exports but with severe performance cuts (e.g., H20 offers only 15% of A100’s power). Chinese firms adopt a "dual-track strategy"—importing while accelerating R&D.

8. Shanghai Securities: LLM Race Fuels Compute Investment

OpenAI’s ChatGPT Agent launch spiked compute demand, while GPT-5 expectations drive infrastructure investment (optical modules, PCBs, liquid cooling, etc.). Domestic GPU makers gain export-substitution orders.

9. China’s Compute Industry Chain Hits New Highs

Sectors like optical modules (↑27.45% weekly), liquid cooling, and IDC led stock gains. Leaders like InnoLight and Eoptolink expanded production. Huawei Ascend ecosystem now hosts 6.65M developers and 8,800 partners.

️ 10. Shanghai Aims to Be Global AI Compute Hub

Shanghai designated AI as a top-priority industry. It leads China in composite compute power. Xuhui’s "Moyuan Space" and Pudong’s "Moli Community" form an east-west innovation chain, advancing chip-to-application R&D.

Weekly Trends

1. Domestic Breakthroughs: Ascend 384 and GritTech G100 mark hardware leaps;

2. LLM Arms Race: GPT-5 and Step-3 drive demand for high-precision compute (e.g., FP8);

3. Edge Computing Rises: NVIDIA Jetson Thor accelerates physical-world AI;

4. Global Chip Dynamics: "Limited thaw" in exports to China parallels domestic substitution.