➀ NVIDIA's GB200 AI servers face supply chain delays; ➁ Mass production may be delayed until Q2 or Q3 2025 due to high design specifications; ➂ The GB200 NVL72 model requires advanced NVLink connectivity and high TDP, prompting a shift towards liquid cooling solutions.
Recent #AI server news in the semiconductor industry
➀ AI servers are expected to consume over 1000 kW of power due to NVIDIA's upcoming Rubin Ultra AI GPU and HBM4 memory; ➁ Current AI servers use 130-250 kW, but future models like Blackwell Ultra and Rubin AI servers will require 250-900 kW, with up to 576 GPUs per rack by 2026-2027; ➂ NVIDIA's future AI servers with Rubin Ultra AI GPUs will use 1000 kW or more by 2028-2029+.
➀ Quanta Computer plans to increase AI server production in the US due to high demand; ➁ NVIDIA's GB200 is being trial-produced in Taiwan and the US, with production ramping up in Q1 2025; ➂ Quanta has production capacity on both the east and west coasts of the United States and will continue to expand its factories.
➀ NVIDIA's Blackwell AI GPUs are 'sold out' for the next 12 months due to high demand; ➁ The GPUs are in high demand among major tech companies like Google, Meta, and Microsoft; ➂ Despite the delay in production, Blackwell is expected to deliver significant performance improvements for AI training and inference.
➀ NVIDIA's new GB200 NVL72 AI server faces significant challenges due to its 132kW TDP; ➁ This makes it the highest-power-consuming server in history; ➂ The development of the server is experiencing delays due to technical issues.
➀ NVIDIA is reportedly halting development of its dual-rack 72-way GB200-based NVL36x2 AI server; ➁ The focus is shifting to single-rack NVL36 and NVL72 AI servers; ➂ The decision is driven by challenges in NVL36 development and customer preferences.