内容来源：https://aibusiness.com/agentic-ai/google-readies-purpose-built-ai-chip

内容总结：

谷歌云近日正式发布第七代自研AI芯片Ironwood TPU，标志着其在高端人工智能算力领域向英伟达发起新一轮挑战。这款专为大规模模型训练、强化学习及高并发推理场景设计的芯片，经过数月测试后将于近期全面上市。

据官方技术博客披露，Ironwood TPU采用突破性互联架构，单个超级模块可集成9,216枚芯片，通过每秒9.6太比特的专有互联网络实现协同工作，配合1.77拍字节的高速共享内存，能有效突破大规模AI运算的数据瓶颈。性能测试显示，其运算速度较第五代TPU提升十倍，比第六代Trillium芯片快四倍。

该芯片的商用化进程已获行业重要伙伴背书。人工智能公司Anthropic宣布将部署百万量级TPU集群，其计算部门负责人詹姆斯·布拉德伯里指出，Ironwood在推理性能与训练扩展性的双重提升，将助力Claude模型在保证服务质量的同时实现高效扩张。虽未披露具体金额，但业内预估该合作规模可达数十亿美元。

同步推出的还有Axion系列Arm架构CPU的升级方案，以及首款基于Arm架构的裸金属实例C4A的预览计划。 Alphabet最新财报显示，AI基础设施需求已成为重要增长引擎，公司首次实现单季度营收突破千亿美元。首席执行官桑达尔·皮查伊在业绩说明会上强调，基于TPU与GPU的解决方案将持续获得强劲市场需求，公司正加大投入以应对算力需求增长。

中文翻译：

由谷歌云赞助
如何选择首个生成式AI应用场景
开展生成式AI项目时，应首先关注能够优化人机信息交互体验的领域。

其第七代Ironwood TPU专为处理计算密集型任务设计，包括模型训练、强化学习、推理及模型服务等环节。

随着人工智能领域竞争日趋白热化，谷歌正瞄准英伟达，推出其迄今最强大的芯片。第七代Ironwood张量处理器（TPU）自四月起进入测试阶段，将于未来数周内全面上市。

据谷歌官方博客介绍，Ironwood芯片"专为最严苛的工作负载量身打造：无论是大规模模型训练、复杂强化学习，还是高吞吐量低延迟的AI推理与模型服务，均能胜任"。谷歌宣称其性能实现显著提升，运算速度较第五代TPU提升十倍，较第六代Trillium芯片快四倍。

其互联设计尤为值得关注：通过将多个芯片连接构成集群，使其能够协同运作形成统一整体。Ironwood TPU最多可将9,216枚芯片连成超级集群，借助谷歌专有的片间互联网络（传输速率达9.6Tb/s）及1.77PB共享高带宽内存，即使应对最复杂的模型也能突破数据瓶颈。

谷歌此次发布时机把握精准。正如其指出，虽然谷歌Gemini、Anthropic的Claude等前沿模型均在TPU上完成训练与部署，但自主AI工作流与推理需求的兴起，要求计算架构与加速机器学习之间实现更紧密协同，这为Ironwood等定制芯片创造了机遇。

事实上，Anthropic已承诺将部署百万量级TPU集群。该公司计算业务负责人詹姆斯·布拉德伯里在博客中表示："随着（对Claude的）需求呈指数级增长，我们在突破AI研究与产品开发边界的同时，正持续扩充算力资源。Ironwood在推理性能与训练扩展性上的双重提升，将助力我们在保障用户期待的响应速度与可靠性的同时，实现高效扩容。"虽未披露具体金额，这笔交易规模可能达数十亿美元。

与Ironwood同步，谷歌还为其通用型Arm架构Axion系列CPU推出增强版本，并宣布即将推出首款Arm架构裸金属实例C4A金属版的预览版本。

不出所料，市场对AI基础设施的旺盛需求正显著影响谷歌及其母公司Alphabet的财务状况——Alphabet最新财报显示，其第三季度营收首次突破千亿美元大关。

Alphabet与谷歌首席执行官桑达尔·皮查伊在业绩说明会上向投资者透露："我们观察到市场对AI基础设施产品（包括TPU与GPU解决方案）存在巨大需求。这不仅是过去一年业绩增长的关键驱动力，展望未来，我们坚信需求仍将保持强劲态势，公司正持续投入以满足市场需求。"

您可能还喜欢

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Its seventh-generation Ironwood TPU was designed to handle compute-intensive tasks such as model training, reinforcement learning, inferencing and model serving.
Google is taking aim at Nvidia and rolling out its most powerful chip yet, as the battle to win AI business continues to heat up.
The seventh-generation Ironwood tensor processing unit (TPU) has been in testing since April and will be generally available in the coming weeks.
Ironwood has been "purpose-built for the most demanding workloads: from large-scale model training and complex reinforcement learning (RL) to high-volume, low-latency AI inference and model serving," according to a Google blog post.
It also constitutes an upgrade in performance, operating ten times faster than the company's fifth-generation TPU and four times quicker than its sixth-gen one, Trillium, the company stated.
Of particular note is the design, with each chip connected to another, creating a pod, which allows them to work as a single unit.
With Ironwood TPUs, up to 9,216 chips can be linked in a superpod, and this connectivity -- via Google's proprietary Inter-Chip Interconnect network operating at 9.6 terabits per second -- is able to overcome data bottlenecks for even the most demanding models, assisted by access to 1.77 petabytes of shared high-bandwidth memory.
Google's release is strategically timed. As it points out, frontier models such as Google Gemini and Anthropic's Claude train and serve on TPUs, but the rise of agentic AI workflows and inference requires greater coordination between compute and accelerated machine learning, creating opportunities for custom silicon such as Ironwood.
Indeed, Anthropic has already committed to accessing up to 1 million TPUs. James Bradbury, the firm's head of compute, said in the blog post: "As demand [for Claude] continues to grow exponentially, we're increasing our compute resources as we push the boundaries of AI research and product development. Ironwood's improvements in both inference performance and training scalability will help us scale efficiently while maintaining the speed and reliability our customers expect."
No financials have been confirmed for the deal, but it is likely to stretch to billions of dollars.
In tandem with Ironwood, Google is debuting improved options for its Axion family of Arm-based CPUs for general purpose workloads, plus it announced that C4A metal, its first Arm-based bare metal instance, will be coming soon in preview.
As might be expected, insatiable demand for AI infrastructure is having a significant effect on the finances at Google and parent company Alphabet, with the latter reporting that Q3 was its first ever quarter where it reached $100 billion in revenue.
Sundar Pichai, Alphabet and Google CEO, told investors during an earnings call: "I would say we are seeing substantial demand for our AI infrastructure products, including TPU-based and GPU-based solutions. It is one of the key drivers of our growth over the past year, and I think on a going-forward basis, I think we continue to see very strong demand, and we are investing to meet that."
You May Also Like