A naukri.com initiative
Cloudblog
2M
156
Image Credit: Cloudblog
Introducing A4X VMs powered by NVIDIA GB200 — now in preview
- The preview of A4X VMs, powered by NVIDIA GB200 NVL72, introduces an integrated system with 72 Blackwell GPUs and 36 Arm-based NVIDIA Grace CPUs connected via fifth-generation NVIDIA NVLink to cater to reasoning models and complex AI tasks.
- A4X VMs offer enhanced training performance with over 1 exaflop per GB200 NVL72 system, allowing a 4X increase in LLM training performance compared to A3 VMs powered by NVIDIA H100 GPUs.
- Scalability and parallelization are key features, facilitating the deployment of models across tens of thousands of Blackwell GPUs and maximizing GPU utilization.
- The optimized A4X architecture is designed for low-latency inference, especially for reasoning models using chain-of-thought techniques, sharing memory and workload across all 72 GPUs for improved performance.
- Google Cloud's infrastructure advantage with A4X VMs includes Hypercompute Cluster for managing large clusters, high-performance networking fabric for efficient GPU-to-GPU traffic, and advanced liquid cooling for peak computational performance.
- The A4X system is integrated with software optimizations for Arm-based hosts, ensuring performance-optimized software like libraries and drivers for popular frameworks such as PyTorch and JAX.
- A4X VMs offer native integration with Google Cloud products and services, including Cloud Storage FUSE for better training throughput, GKE compatibility for maximized resource usage, and Vertex AI Platform for accelerated AI project development.
- Customers like Magic have chosen A4X VMs for their AI supercomputing needs, leveraging enhanced performance and scalability for demanding workloads.
- The collaboration between Google and NVIDIA will bring NVIDIA DGX Cloud, a managed AI platform, to A4X VMs, offering improved performance and scalability for AI initiatives.
- A guide to choosing between A4 and A4X VMs is provided, with A4X recommended for complex reasoning models and large-scale workloads, while A4 is suitable for diverse AI model architectures and workloads.
- For more information about A4X, interested parties are advised to contact their Google Cloud representative and explore its benefits at Google Cloud Next.
Read Full Article
9 Likes
For uninterrupted reading, download the app