英伟达不再独霸？谷歌AI芯片算力追平B200

Core Insights - Google introduced its seventh-generation TPU, Ironwood, at the Google Cloud Next 25 conference, marking it as the most powerful TPU to date designed for large-scale AI reasoning models [1][2] TPU Overview - TPU (Tensor Processing Unit) is a specialized AI chip designed to accelerate deep learning tasks, first introduced by Google in 2015, with the first generation released in 2016 [2] - Ironwood represents a shift in AI infrastructure from reactive models providing real-time information to proactive models generating insights and interpretations [2] Technical Specifications - Ironwood can support a maximum cluster of 9,216 liquid-cooled chips, achieving a peak performance of 42.5 ExaFlops, equating to 42.5 quintillion operations per second [2] - The chip supports FP8 computation, with a performance of 4,614 TFlops, slightly surpassing NVIDIA's B200 at 4,500 TFlops, and has a memory bandwidth of 7.2 TBps, which is lower than B200's 8 TBps [3] - Ironwood features the third-generation SparseCore accelerator, designed to enhance financial and scientific computations, initially aimed at accelerating recommendation models [3] Comparative Analysis - Comparison of TPU generations shows significant advancements: - Pod Size: Ironwood (9,216 chips) vs. TPU v4 (4,896 chips) and TPU v5p (8,960 chips) - HBM Bandwidth: Ironwood (192 GB) vs. TPU v4 (32 GB) and TPU v5p (95 GB) - Capacity: Ironwood (7.4 TBs HBM) vs. TPU v4 (1.2 TBs HBM) and TPU v5p (2.8 TBs HBM) - Peak performance per chip: Ironwood (4,614 TFLOPS) vs. TPU v4 (275 TFLOPS) and TPU v5p (459 TFLOPS) [4] - Ironwood's performance per watt is double that of the previous generation TPU, Trillium, and its chip capacity is six times larger, allowing for handling of larger models and datasets [4] Future Plans - Google plans to integrate TPU v7 into its cloud AI supercomputing services, which will include recommendation algorithms, Gemini models, and AlphaFold [4] - OpenAI's co-founder Ilya Stutskever's AI startup, Safe Superintelligence, is utilizing Google Cloud's TPU chips for its AI research [5]