GCP Professional Cloud Architect Practice Question

Your enterprise plans to train a 500-billion-parameter language model on Google Cloud. The data science team wants to iterate quickly, while finance requires that the infrastructure cost scales down to zero when the training cluster is idle. You must design the compute environment so that it

provides a high-bandwidth, low-latency fabric between thousands of accelerators,
lets the team choose between NVIDIA H100 GPUs today and future TPU versions without re-architecting, and
exposes simple hooks so Vertex AI Pipelines can create and tear down the cluster on demand. Which Google Cloud capability most directly satisfies these requirements and should be central to your design?

Leverage Google Cloud's AI Hypercomputer platform accessed through Vertex AI custom training jobs.
Create an unmanaged MIG of A3 UltraGPU VMs and script scale-down using Cloud Functions and Cloud Scheduler.
Deploy a GKE Autopilot cluster with node pools that use Spot A3 GPU VMs and rely on Cluster Autoscaler for scale-to-zero.
Reserve a dedicated TPU v5p pod via Compute Engine and attach it to Vertex AI using custom containers.

Report Issue

Answer Description

AI Hypercomputer is Google Cloud's integrated supercomputing architecture that combines cutting-edge hardware (such as A3 instances with NVIDIA H100 GPUs and the latest TPU v5p pods), custom high-performance networking (Jupiter fabric), and a fully managed software stack that is exposed through services like Vertex AI. Because resources are provisioned through standard Google Cloud APIs, the training cluster can be instantiated by Vertex AI Pipelines when a run starts and automatically released when it finishes, so costs drop to zero when idle. Other options provide parts of the solution-individual accelerator VMs, GKE, or manual reservation management-but none deliver the turnkey, fabric-optimized, multi-accelerator platform with flexible consumption models required for large-scale model training.

Ask Bash

Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.