COMPUTEX 2025

Core Insights - The article discusses the advancements in AI technology as presented by NVIDIA CEO Jensen Huang at COMPUTEX 2025, highlighting the evolution of AI from perception to reasoning and physical AI [1] AI Evolution Path - The evolution of AI is categorized into four stages: 1. Perception AI: Understanding patterns like speech and image recognition 2. Generative AI: Transitioning from understanding to generating content across multiple modalities 3. Reasoning AI: Focusing on complex reasoning capabilities, utilizing techniques like "Chain of Thought" and "Tree of Thought" 4. Physical AI: Understanding physical concepts such as inertia and causality, crucial for the next AI era [1] GB300 and Blackwell Architecture - The GB300 system, based on the new Grace Blackwell architecture, has been in production since early this year, with significant upgrades including a 1.5x increase in inference performance and a 2x increase in network capability [2] - The system features 100% liquid cooling and maintains the same physical footprint as previous models, with a single node performance of approximately 40 petaflops [2] NVLink and CoWoS-L Technology - NVIDIA has developed a new collaborative process with TSMC called CoWoS-L to create larger chips, enhancing performance through NV-Link technology, which offers a data transfer speed of 7.2TB/s [3] - The NV-Link architecture connects multiple GPUs within a single rack, achieving a bandwidth of 130 terabytes/s, necessitating liquid cooling due to high power requirements [3] NVLink Fusion - NVLink Fusion is introduced to allow partners to build semi-custom AI infrastructure solutions, enabling integration of custom ASICs into NVIDIA's ecosystem [4] - This technology facilitates the mixing of NVIDIA components with partner-specific chips, enhancing the flexibility of AI infrastructure [5] DGX Spark and Workstations - DGX Spark has entered full production, designed for AI-native developers, offering 1 petaflops of computing power and 128GB of memory for prototyping and early development [6] - NVIDIA also launched desktop-level DGX supercomputers, capable of running AI models with up to 1 trillion parameters, suitable for home use [6] Enterprise AI Solutions - The RTX Pro Enterprise server integrates x86 architecture and supports various AI agents, showing significant performance improvements over previous models [7]