Global GPU Computing News【20250612】

1. VeriSilicon Launches High-Performance GPGPU-AI Computing IP
   VeriSilicon introduced its next-gen GPGPU-AI computing IP, supporting mixed-precision computing (INT4/FP4) and 3D-stacked memory integration. Optimized for automotive electronics and edge servers, it efficiently handles inference tasks for large models like DeepSeek.

2. Inbobox Launches Elastic GPU Container Service
   Inbobox Cloud unveiled its GPU computing power product suite, introducing a "Unit Effective Computing Power Cost" standard. It enables minute-level billing and elastic resource allocation, significantly reducing enterprise inference costs.

3. Moore Threads Deploys Thousand-Card Clusters in Multiple Cities
   Moore Threads’ KUAE computing cluster solution completed deployments in Beijing’s Yizhuang/Miyun and Nanjing. It supports 200B-parameter model training (e.g., Aquila2) with 91% thousand-card scaling efficiency.

4. JD Cloud’s vGPU Platform Boosts Financial Sector Efficiency
   China Asset Management adopted JD Cloud’s vGPU platform, achieving 70% higher GPU utilization and 98% resource fulfillment for inference tasks, driving cost efficiency in finance.

5. National Data Administration Accelerates Computing Network
   China’s integrated computing power network will launch nationwide in 2025, addressing regional resource imbalances and targeting 50% total computing power growth.

6. Nvidia GB200 Shipment Forecast Downgraded
   Morgan Stanley reports GB200 chip shipments may drop to 25k units (from 35k) in 2025 due to domestic substitution, potentially costing supply chains $35B.

7. Meta’s 24K GPU Cluster Powers Llama 3 Training
   Meta’s custom 24K-GPU cluster trained Llama 3 with >400 TFLOPS per-card utilization, highlighting 10k+ GPU clusters as the new standard for large model training.

8. Edge AI Drives GPGPU Architecture Upgrades
   VeriSilicon’s team notes surging demand for Transformer support in edge servers, urging next-gen GPGPUs to integrate sparse computing and inter-core optimization.

9. Nvidia H100 NVL Chip Expands Commercial Use
   The dual-GPU NVLink H100 NVL chip accelerates deployment, delivering 10x higher inference performance than A100 for trillion-parameter models.

10. Multimodal Computing Costs Challenge AIGC Adoption
    Video generation (e.g., Sora) and multimodal apps push computing demands, driving enterprises toward hybrid heterogeneous deployments to balance cost and performance.

Summary

• Domestically: Accelerated self-reliance via local cluster deployments (Moore Threads), computing optimization (JD Cloud), and policy support (national computing network);

• Internationally: Intensified high-performance chip competition, edge/multimodal-driven architecture upgrades, and heightened supply chain risks (e.g., Nvidia’s shipment cuts).