半导体行业观察·2025-10-15 02:48Core Insights - Oracle and AMD are expanding their long-term partnership to enhance AI capabilities, with Oracle's cloud infrastructure set to deploy a significant AI supercluster powered by AMD's Instinct MI450 series GPUs starting in Q3 2026 [9][10]. Group 1: AI Supercluster Details - The AI supercluster will initially deploy 50,000 Altair GPU slots, with plans for gradual expansion in 2027 and beyond [6][9]. - The MI450 GPU is expected to deliver approximately 40 petaflops of peak computing power with 432 GB of HBM4 memory and a memory bandwidth of 19.6 TB/s [2][11]. - The Helios rack design will support up to 128 MI450X GPUs, providing a total of 1.45 exaflops at FP8 precision and 2.9 exaflops at FP4 precision [5][10]. Group 2: Technical Specifications - The MI450 series will utilize TSMC's 2nm process technology, enhancing performance and efficiency [2]. - Each Helios rack will feature advanced networking capabilities, including UALink for GPU interconnectivity and Pensando DPU technology for enhanced security and performance [12][14]. - The architecture is designed to optimize performance density, cost, and energy efficiency, crucial for large-scale AI operations [12][13]. Group 3: Market Context and Future Outlook - The demand for large-scale AI capacity is accelerating, necessitating flexible and open computing solutions [10]. - Oracle's Acceleron network architecture aims to reduce equipment layers in large-scale AI deployments, indicating a strategic move towards more integrated solutions [7][11]. - The anticipated total cost for the deployment of 700 racks is estimated to be between $3.5 billion and $4 billion, reflecting the high demand and limited supply of GPUs [6][7].