Core Insights - AMD has announced a significant agreement with Oracle for 50,000 GPUs and showcased its upcoming super AI rack planned for the second half of 2026 [1][33] - The AMD Helios rack has notable design differences compared to Meta's version, particularly in power distribution and network layout [33] Group 1: AMD Helios Rack Features - The AMD Helios rack features a management switch and power supply at the top, with a power shelf containing compute nodes below [3][5] - The design includes a robust handle on the switch tray, indicating a focus on durability [9] - The rack utilizes EDSFF E1.S SSDs and has transitioned from 2.5-inch U.2 connectors in PCIe Gen6 [10] Group 2: Comparison with Meta's Rack - Meta's rack design differs significantly, with Ethernet switches placed at the top instead of power supplies, indicating a major architectural change [15][16] - The internal arrangement of power and network components in Meta's rack allows for a more efficient power delivery system [29] - Both racks share similarities in width and the sandwich layout of compute and switch trays, but the internal configurations are distinct [33] Group 3: Performance Metrics - AMD's Helios AI rack is equipped with the MI400 series, boasting 72 GPU domains, 260 TB/s scale-up bandwidth, and 2.9 EF FP4 and 1.4 EF FP8 FLOPS [35] - The HBM4 memory capacity is 31 TB with a memory bandwidth of 1.4 PB/s, showcasing AMD's competitive edge in high-performance computing [35] - The Helios AI rack has attracted significant investment interest from major AI companies [36]