Original post · AI translation
Google 的 TPU 将首次在 Google 以外的超大规模云平台上运行，Oracle 将从 2027 日历年开始部署算力，并在随后数年逐步扩大规模。
Oracle 提供电力、冷却和数据中心网络，其 Acceleron 网络将 TPU 集群连接到 Oracle 数据库和存储系统，使企业无需跨云传输，就能在其数据记录旁微调和提供模型服务。
Google 声称，其第 8 代 TPU 每颗芯片可提供最高 12.6 PFLOPS 的 FP4 算力，在大规模训练任务中，集群超过 97% 的时间都在进行有效工作，并且在低延迟模型服务方面，每美元性能最高可提升 80%。
Google 表示，其开源 TorchTPU 技术栈无需重写模型，就能让现有 PyTorch 代码以及 vLLM 和 SGLang 等工具在 TPU 上运行，从而降低将工作负载从 NVIDIA GPU 迁出的成本。
这笔交易将 Google 向自家云平台之外销售 TPU 的努力推进到一家直接竞争的超大规模云服务商；这项努力始于今年早些时候向规模较小的服务商销售 TPU。
Original
Google’s TPUs will run on a hyperscale cloud outside Google for the first time, with Oracle deploying capacity from calendar 2027 and ramping it over several years. Oracle supplies the power, cooling and data center networking, and its Acceleron network links TPU clusters to Oracle databases and storage, so enterprises can fine-tune and serve models beside their records without cross-cloud transfers. Google claims its 8th-generation TPUs deliver up to 12.6 PFLOPS of FP4 compute per chip, keep clusters doing productive work over 97% of the time on large training runs, and give up to 80% better performance per dollar for low-latency serving. Google says its open-source TorchTPU stack runs existing PyTorch code, along with tools such as vLLM and SGLang, on TPUs without model rewrites, which lowers the cost of moving work off NVIDIA GPUs. The deal carries Google’s effort to sell TPUs beyond its own cloud, which began earlier this year with sales to smaller providers, into a direct hyperscale competitor.