解构亚马逊最强芯片，GPU迎来劲敌

Core Insights - The article discusses the anticipation surrounding AWS's Trainium4 XPU, which is expected to be delivered by late 2026 or early 2027, causing concerns among users currently waiting for Trainium3 [1][18] - Trainium3 is highlighted as a significant improvement over its predecessors, offering enhanced performance and efficiency, but Trainium4 is projected to bring even greater advancements [1][4] Summary of Trainium3 Specifications - Trainium3 utilizes TSMC's 3nm process technology, providing double the computing power and a 40% increase in energy efficiency compared to previous models [4][6] - The UltraServer configuration for Trainium3 can support up to 64 slots, with a total HBM memory bandwidth that is 3.9 times greater than Trainium2 [6][14] Performance Metrics - Trainium3 UltraServer shows a 4.4 times increase in overall computing power compared to Trainium2 UltraServer, with a significant increase in token output per megawatt [6][8] - The architecture includes five types of computing units, enhancing its capability for high-performance computing and AI workloads [9][10] Future Prospects with Trainium4 - Trainium4 is expected to support a new architecture, NeuronCore-v5, which will include native FP4 support, potentially increasing performance by six times compared to Trainium3 [18][21] - The anticipated HBM memory capacity for Trainium4 is projected to be double that of Trainium3, with bandwidth expected to quadruple [18][21] Architectural Improvements - Trainium4 is speculated to incorporate both NVLink and UALink ports, allowing for enhanced connectivity and performance [19][20] - The design aims to balance computation, memory, and interconnect performance, with a potential increase in core count to achieve higher efficiency [20][21]