NVIDIA and Google Cloud announced a new phase of their decade-long partnership at Google Cloud Next in Las Vegas. The goal of the collaboration is to accelerate the development of agentic and physical artificial intelligence, expanding the capabilities of the Google Cloud AI Hypercomputer platform for AI factories. This will allow the translation of laboratory developments into production solutions, implementing AI agents for managing complex processes, as well as robots and digital twins in production.
One of the key announcements was the new A5X serverless instances, built on the NVIDIA Vera Rubin architecture. They significantly increase efficiency, providing up to 10 times lower inference cost per token and 10 times higher token throughput per megawatt compared to previous-generation solutions. These systems integrate NVIDIA ConnectX-9 SuperNICs and Google Virgo networking technology, allowing them to scale up to 80,000 NVIDIA Rubin GPUs in a single cluster and up to 960,000 GPUs in a multi-site cluster.
In addition, preview versions of Google Gemini models for Google Distributed Cloud were presented, which will run on NVIDIA Blackwell and NVIDIA Blackwell Ultra GPUs. This innovation allows customers to securely use Google's advanced AI models in their own environments where the most sensitive data is processed. The platform also supports agentic AI based on the Gemini Enterprise Agent Platform, using open NVIDIA Nemotron models and the NVIDIA NeMo framework.
The Google Cloud portfolio now includes an expanded range of NVIDIA Blackwell-based solutions. These include A4 virtual machines with NVIDIA HGX B200 systems, A4X rack-mount virtual machines with NVIDIA GB200 NVL72, A4X Max with NVIDIA GB300 NVL72, as well as fractional G4 virtual machines with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. This flexibility allows customers to scale compute resources from one-eighth of a GPU to thousands of Blackwell GPUs in clusters with NVIDIA NVLink 5 technology, adapting them to various needs.
The new comprehensive platform is optimized for a wide range of workloads: from Mixture-of-Experts reasoning and multimodal inference to data processing and complex simulations for physical AI and robotics. Leading AI labs, including Thinking Machines Lab and OpenAI, are already using this infrastructure to scale applications, train models, and perform large-scale inference, including for ChatGPT. Special attention is paid to data confidentiality. Thanks to NVIDIA Confidential Computing on the NVIDIA Blackwell platform, Gemini models can function in a protected environment. This ensures the encryption of requests and data for fine-tuning, making them inaccessible to unauthorized parties, including infrastructure operators. Data protection in the public cloud is also provided by the preview version of confidential G4 VMs with NVIDIA RTX PRO 6000.
Mark Lohmeyer, Vice President of Google Cloud, noted: "The next decade in AI will be defined by customers' ability to run the most demanding workloads on an integrated and AI-optimized infrastructure stack." This partnership provides customers with unprecedented flexibility in training, customizing, and serving models, significantly improving performance, reducing costs, and ensuring the sustainability of solutions.
Sources
Replies (0)
No replies in this topic yet.
