aifollow.news Search
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time SelectedAI score67

Oracle to deploy Google TPUs on its cloud starting in 2027Machine translation

Automatically verified and published · Generated and evidence-checked automatically; not reviewed by a human.

AI introduction

Oracle will provide power, cooling and data center networking, while its Acceleron network connects TPU clusters to Oracle databases and storage. This would let enterprises fine-tune and serve models alongside their records without cross-cloud transfers. Google claims its eighth-generation TPUs offer up to 80% better performance per dollar for low-latency serving; Oracle will deploy capacity from 2027 and ramp it up over several years.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

Google 的 TPU 将首次在 Google 之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,并在此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,使企业能够在自身数据记录旁微调和提供模型服务,而无需跨云传输。

Google 声称,其第 8 代 TPU 每颗芯片可提供高达 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 软件栈可让现有 PyTorch 代码以及 vLLM 和 SGLang 等工具在 TPU 上运行,无需重写模型,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 的努力扩展到了一个直接与其竞争的超大规模云服务商;这一努力始于今年早些时候向较小规模的服务商销售 TPU。

Source: Rohan Paul · x.com

Industry
Found an error? Send a correction