Nvidia's Upcoming Blackwell Platform Set to Launch on Google Cloud in Early 2025

Google Cloud Next Unveils Exciting Innovations in Las Vegas

This week marks Google Cloud Next in Las Vegas, showcasing a wave of new instance types and accelerators on the Google Cloud Platform. Highlighting this year’s announcements are advanced AI accelerators, including Google’s custom Arm-based Axion chips and innovations from Nvidia.

Just a few weeks ago, Nvidia unveiled its Blackwell platform, yet Google’s integration of these machines is not expected until 2025. Anticipated support for Nvidia’s high-performance HGX B200, designed for AI and HPC workloads, along with the GB200 NBL72 for large language model (LLM) training, is set to launch in early 2025. A notable detail from Google’s announcement is that the GB200 servers will utilize liquid cooling for improved efficiency.

While this might appear as an early preview, Nvidia announced that its Blackwell chips won’t be publicly available until the end of 2023.

Enhancing AI Workloads with A3 Mega Instances

For developers seeking immediate power to train LLMs, Google has introduced the A3 Mega instance, developed in collaboration with Nvidia. This instance features cutting-edge H100 GPUs paired with an innovative networking system that offers up to double the bandwidth per GPU compared to previous models.

Moreover, Google is launching the A3 confidential instance, designed to enhance data confidentiality and integrity during sensitive data use and AI workloads training. This builds on Google’s existing confidential computing services, which encrypt data during use, ensuring that data transfers between Intel’s CPU and the Nvidia H100 GPU are secure through protected PCIe. Importantly, no code modifications are needed, according to Google.

Introducing Cloud TPU v5p Processors

In a significant achievement, Google has rolled out its Cloud TPU v5p processors—the most advanced of its in-house AI accelerators to date. These processors boast a twofold increase in floating-point operations per second and tripling memory bandwidth speed, positioning users for greater computational efficiency.

Supporting Infrastructure with AI-Optimized Storage

To complement its powerful chips, Google has revealed new AI-optimized storage solutions. The next-generation Hyperdisk ML, currently in preview, aims to enhance model loading speeds by up to 3.7 times, boosting overall performance for AI applications.

Additionally, Google Cloud is launching traditional instances powered by Intel’s fourth- and fifth-generation Xeon processors. The new general-purpose C4 and N4 instances will feature the fifth-generation Emerald Rapids Xeons, with the C4 emphasizing performance and the N4 focusing on cost-effectiveness. The C4 instances are in private preview, while the N4 machines are available now.

The new offerings also include the C3 bare-metal machines, utilizing previous-generation Intel Xeons, and the X4 memory-optimized bare metal instances, both in preview. Furthermore, the Z3 instance is Google Cloud’s inaugural storage-optimized virtual machine, designed to provide unparalleled IOPS among storage-optimized instances across leading cloud services.

Stay tuned for more updates from Nvidia's keynote at GTC, which offered some unexpected announcements!