Google Cloud Unveils A4 VMs with NVIDIA Blackwell GPUs for AI

Originally published at https://ssojet.com/blog/google-cloud-unveils-a4-vms-with-nvidia-blackwell-gpus-for-ai

Google Cloud has launched its new A4 virtual machines (VMs) powered by NVIDIA's Blackwell B200 GPUs, aimed at enhancing AI workloads significantly. The A4 VM features eight Blackwell GPUs interconnected via fifth-generation NVIDIA NVLink, resulting in a 2.25x increase in peak compute capabilities and high bandwidth memory (HBM) capacity over the previous A3 VMs. This performance enhancement is crucial for training, fine-tuning, and inference of complex AI models.

Key features include:

• Enhanced networking with Google's Titanium ML network adapter and NVIDIA ConnectX-7 NICs, enabling 3.2 Tbps of GPU-to-GPU traffic with RDMA over Converged Ethernet (RoCE).
• Integration with Google Kubernetes Engine (GKE), which supports up to 65,000 nodes per cluster, facilitating robust AI infrastructure.
• Access through Vertex AI, Google's unified AI development platform.
• Collaboration with NVIDIA to optimize frameworks such as JAX and XLA for GPU efficiency.

The A4 VMs also feature a hypercompute cluster system that simplifies the deployment and management of large-scale AI workloads, focusing on performance, reliability, and observability.

NVIDIA Blackwell Generally Available in the Cloud

CoreWeave has launched the first NVIDIA GB200 NVL72-based instances, making the NVIDIA Blackwell platform generally available. This setup is designed to support the demands of AI reasoning models, which require extensive computational power and optimized software.

The GB200 NVL72 features:

• Rack-scale NVIDIA NVLink across 72 NVIDIA Blackwell GPUs and 36 NVIDIA Grace CPUs.
• Scaling capabilities to up to 110,000 GPUs powered by NVIDIA Quantum-2 InfiniBand networking.
• Full-stack AI platform integration, including NVIDIA Blueprints for customizable workflows and NVIDIA NeMo for model training.

CoreWeave's instances offer 400 Gb/s bandwidth per GPU, ensuring high-speed communication required for deploying AI agents.

A3 Ultra VMs with NVIDIA H200 GPUs

The A3 Ultra VMs powered by NVIDIA H200 GPUs have become generally available. These VMs are part of Google Cloud's AI Hypercomputer, designed for running distributed workloads efficiently.

Key features include:

• Enhanced GPU-to-GPU network bandwidth and HBM, delivering optimized performance for AI workloads.
• Integration with Google Kubernetes Engine, allowing seamless deployment of VMs.
• Support for high-performance computing (HPC) capabilities.

The A3 Ultra VMs are built on the latest Titanium ML network adapter, enabling non-blocking GPU-to-GPU communication with RDMA over Converged Ethernet.

Nvidia's Project DIGITS Desktop AI Box

At the 2025 CES event, Nvidia unveiled a new desktop computer, Project DIGITS, developed in collaboration with MediaTek. This system is powered by a Grace CPU and Blackwell GPU Superchip, designed for AI and HPC markets.

Key specifications include:

• A GB10 Superchip with 20 Arm cores, aiming for a petaflop of GPU-AI computing performance.
• 128GB of unified memory between CPU and GPU, overcoming limitations of lower-cost consumer video cards.
• Up to 4TB of NVMe storage, designed for efficient operation without extensive power requirements.

Nvidia emphasizes that developers can run large language models locally, enabling a seamless transition to deploying these models on larger cloud infrastructures.

For enterprise clients looking to enhance security and user management, SSOJet provides an API-first platform featuring single sign-on (SSO), multi-factor authentication (MFA), and directory synchronization. Explore our services or contact us at SSOJet to implement secure user management solutions tailored to your needs. Visit SSOJet for more information.