Google Cloud 集成 Nvidia Blackwell 平台以增强人工智能和计算能力

Nvidia Blackwell platform was intended to integrate its advanced future into the Google Computer System Platform, which Google has promised. Initially anticipated for the beginning of 2025, the platform comes with the ultrafast Nvidia HGX B200, which is developed for AI and high-performance computing (HPC) domains as well as the NBL72 GB200 created for the training of large language models (LLMs). Of all the improvements Google will be introducing the first one is liquid cooling to increase the performance and efficiency. This is the first time liquid cooling will be implemented in Google’s cloud infrastructure.

Nvidia Blackwell 平台旨在将其先进的未来集成到谷歌计算机系统平台中，这是谷歌所承诺的。该平台最初预计在 2025 年初推出，配备了专为人工智能和高性能计算 (HPC) 领域开发的超快 Nvidia HGX B200，以及为训练大型语言模型 (LLM) 而创建的 NBL72 GB200。在谷歌将推出的所有改进中，第一项是液体冷却，以提高性能和效率。这是谷歌云基础设施首次采用液体冷却。

Expansion of AI accelerators and computing services

Google introduced new instance types and various accelerators to fulfill every customer’s needs. One of the innovations that Google Cloud has introduced is the custom-made AI processor chip, called the Arm-based Axion, in addition to a wide range of AI accelerators and even from Nvidia that will ultimately create an array in terms of Google’s offerings. Along with the new A3 Mega instance, which is set to be jointly developed with Nvidia, there were other new announcements. 

人工智能加速器和计算服务的扩展谷歌推出了新的实例类型和各种加速器来满足每个客户的需求。谷歌云推出的创新之一是定制的人工智能处理器芯片，称为基于 Arm 的 Axion，此外还有各种人工智能加速器，甚至来自 Nvidia 的加速器，最终将根据谷歌的产品创建一个阵列。除了与 Nvidia 联合开发的新 A3 Mega 实例外，还有其他新公告。

This particular solution is designed with the H100 GPUs which are used in industry-standard nets and a new networking system that can double the bandwidth per each GPU. This is an excellent feature for the developers since it now has the necessary power for training deep LLMs at a high-efficiency rate. 

该特殊解决方案采用行业标准网络中使用的 H100 GPU 和新的网络系统设计，可将每个 GPU 的带宽加倍。这对于开发人员来说是一个出色的功能，因为它现在拥有以高效率培训深度法学硕士的必要能力。

Google also launched its A3 confidential instance, which is designed to be resilient to data breaches and maintain the confidentiality and integrity of sensitive data and AI workloads by encrypting data end-to-end during its transfer and movement between Intel’s CPU and the Nvidia H100 GPU without any need for code changes.

谷歌还推出了 A3 机密实例，该实例旨在抵御数据泄露，并通过在英特尔 CPU 和 Nvidia H100 GPU 之间传输和移动期间对数据进行端到端加密来维护敏感数据和人工智能工作负载的机密性和完整性无需更改任何代码。

Advancements in Google’s own AI accelerators

Google Humanizes: Google is still not afraid to push technical boundaries with its onsite AI supercomputers, Cloud TPU v5 processors are now pasteurized for long-time use by everyone. The processors, which Google claimed are the most advanced AI accelerator they ever had, will be 2x faster in FLOPS than the previously observed and amazing 3x memory bandwidth speed increase. 

谷歌自己的人工智能加速器的进步Google Humanizes：谷歌仍然不害怕通过其现场人工智能超级计算机突破技术界限，Cloud TPU v5处理器现在经过巴氏灭菌，可供每个人长期使用。谷歌声称这些处理器是他们有史以来最先进的人工智能加速器，其浮点运算速度将比之前观察到的惊人的 3 倍内存带宽速度提高 2 倍。

Google came up with AI-optimized storage options like Hyper disk (In preview). This next-generation block storage service offers a way to drastically reduce model loading time (up to 3.7 times according to Google), Google Artisan declares. While the widespread utilization of AI and machine learning remains its core asset, Google Cloud is also increasingly venturing into general infrastructure categories. 

谷歌提出了人工智能优化的存储选项，例如超级磁盘（预览版）。 Google Artisan 声称，这种下一代块存储服务提供了一种大幅缩短模型加载时间的方法（根据 Google 的说法最多可缩短 3.7 倍）。虽然人工智能和机器学习的广泛使用仍然是其核心资产，但谷歌云也越来越多地涉足通用基础设施类别。

The vendor deployed new themes referencing Intel’s fourth- and fifth-generation Xeon CPUs. There is a new general-purpose Z4, Z5, N4, and C4 instances, which will have the fifth-generation Emerald Rapids Xeons. The Z4 is performance-oriented, while N4 offers the best cost-effective solution. Today demonstration version of the C4 instances is in private mode, but N4 will be available for use to all at the current moment.

该供应商部署了引用英特尔第四代和第五代 Xeon CPU 的新主题。新的通用 Z4、Z5、N4 和 C4 实例将配备第五代 Emerald Rapids Xeon。 Z4以性能为导向，而N4则提供最佳性价比的解决方案。今天，C4 实例的演示版本处于私有模式，但 N4 目前可供所有人使用。

Expanding Google cloud services

Google demoed new C3 physical instances, running on the older Intel Xeon processors of the 4th generation, and X4 memory-optimized physical instances, broadening the service options of Google’s infrastructure. The Z-series introduces the Z3 to Google Cloud’s market, being all storage-optimized virtual machines, which lead their class regarding IOPS with no other major cloud service provider.

扩展谷歌云服务谷歌演示了新的C3物理实例，运行在旧的第四代英特尔至强处理器上，以及X4内存优化物理实例，扩大了谷歌基础设施的服务选项。 Z 系列将 Z3 引入 Google Cloud 市场，它是所有存储优化的虚拟机，在 IOPS 方面领先同类产品，没有其他主要云服务提供商。

The announcements made by Google Cloud at the Las Vegas event consolidate the company’s position in the industry. These announcements prove that the industry leader in cloud computing is still going strong and will stay there for good. In 2025, Nvidia will incorporate its Blackwell platform, with processing units, storage solutions and AI accelerators being already active. This puts Google Cloud at the forefront of cloud computing space. 

谷歌云在拉斯维加斯活动上发布的公告巩固了该公司在行业中的地位。这些公告证明云计算领域的行业领导者仍然强劲并将永远保持下去。 2025 年，Nvidia 将整合其 Blackwell 平台，其中处理单元、存储解决方案和 AI 加速器已经投入使用。这使得谷歌云处于云计算领域的前沿。

The transformations not only ideally improve and enlarge Google’s cloud computing sphere, but also serve developers and businesses, as sources for infinite AI/ML, as well as general processing computations. However, with Google Cloud chasing this dynamic cloud computing ecosystem and maintaining its ingenuity, the company’s strategies and inventions represent its strong intention to a more effective, powerful, and diverse computing system in the future.

这些转变不仅理想地改善和扩大了谷歌的云计算领域，而且还为开发人员和企业提供服务，作为无限人工智能/机器学习以及通用处理计算的来源。然而，随着谷歌云追逐这个充满活力的云计算生态系统并保持其独创性，该公司的战略和发明代表了其对未来更有效、更强大、更多样化的计算系统的强烈意图。

Original Story From techcrunch

来自 techcrunch 的原创故事