电子：谷歌Ironwood构建AI推理工厂，与Blackwell形成差异化竞争

Investment Rating - The report does not explicitly provide an investment rating for the industry or specific companies involved Core Insights - Google has launched its 7th-generation Tensor Processing Unit (TPU v7), codenamed "Ironwood," which offers up to 2x improvements in performance and energy efficiency compared to its predecessor, Trillium [6][7] - Ironwood is positioned as one of the fastest AI inference platforms globally, optimized for large language models (LLMs) and mixture-of-experts (MoE) architectures, with a peak performance of 42.5 ExaFLOPS [7][8] - Google is leveraging a "hardware + cloud service" bundling strategy to enhance the appeal of its cloud services, as Ironwood is only available through Google Cloud, which may drive growth in cloud business amid increasing demand for inference capabilities [9][10] Summary by Sections Section 1: Product Launch and Performance - Google officially unveiled TPU v7 (Ironwood) on April 9, 2025, at the Cloud Next conference, achieving significant performance and energy efficiency improvements [6] - Ironwood can support up to 9,216 TPU chips, providing a peak performance of 42.5 ExaFLOPS, making it one of the fastest AI inference platforms [7] Section 2: Competitive Positioning - Google is intensifying its competition with NVIDIA in the AI chip market, but Ironwood is not sold externally and is only available through Google Cloud [8] - The industry trend is shifting towards inference capabilities, as seen with the release of Llama 4, which focuses on inference performance [8] Section 3: Comparison with NVIDIA - A comparison between Google TPU v7 and NVIDIA's Blackwell architecture shows that TPU v7 focuses on inference while Blackwell targets general-purpose AI acceleration [10][11] - TPU v7 achieves a peak performance of 42.5 ExaFLOPS, while NVIDIA's GB200 card reaches 20 PFLOPS, indicating TPU's advantage in aggregate compute but NVIDIA's lead in precision flexibility and commercial availability [11]