SemiAnalysis 报道了 Google 张量处理单元 (TPU) 在人工智能模型训练市场份额的增长，重点介绍了 Anthropic 等外部实体为其 Claude 模型发布日益增长的采用。该分析详细介绍了即将推出的 TPU 架构，包括具有用于大规模扩展和降低延迟的 BroadFly 拓扑的 TPUv8i，以及利用不同环面配置以适应不同 Pod 大小的 TPUv7 Ironwood 变体。TPU 在 Google 内部使用之外的可用性扩展为人工智能实验室在训练硬件市场提供了竞争性选择。 AI
影响 Google 的 TPU 正在成为人工智能训练更可行的替代方案，可能会影响主要人工智能实验室的硬件选择。
排序理由 对人工智能加速器的硬件供应链和采用趋势的分析。
AI 生成摘要 · Google Gemini · 来自 4 个来源。 我们如何撰写摘要 →
Our AI Accelerator Supply Chain model contains the latest updates on the massive traction in external TPU sales and rentals, as well as the implications for the supply chain. 4/4🧵
With the upcoming TPUv8i BroadFly topology, it can scale up to 1,024 chips within a 7-hop radius, compared with the traditional 3D torus 16-hop radius. This enables lower latency, making it easier to overlap communication with compute without incurring the cost of a dedicated htt…
TPUv7 Ironwood comes in two variants: TPUv7x, which uses a 3D torus for physical pod sizes of up to 9,216 chips, and TPUv7, which uses a 256-chip 2D torus. 2/4🧵 https://t.co/BE2aSz8wg5
Successfully training models on TPUs has been demonstrated by Anthropic through the past five-plus successful Claude releases. This is positive for the ML community, as Google TPUs continue to gain market share outside of internal Google workloads, giving frontier AI labs a https…