An AI research team is deploying multi-node large language model (LLM) fine-tuning jobs that take between 3 and 5 days to complete. The workload has the following architectural requirements: The GPU compute instances must be densely allocated to minimize inter-node network latency during distributed training. The jobs are not fault-tolerant and cannot handle mid-execution preemption or interruptions. Start times are flexible, allowing jobs to queue until the required GPU capacity is fully available. The organization wants to optimize costs with discounted pricing without paying for reserved capacity during idle periods. Which AI Hypercomputer consumption option should you recommend?

Flex-start using Dynamic Workload Scheduler: ### What Flex-start is **Flex-start** is an AI Hypercomputer and Google Cloud compute consumption model powered by the **Dynamic Workload Scheduler (DWS)**. It allows teams to request dense compute capacity for short-duration workloads lasting up to seven days without committing to a long-term reservation. ### How it meets the requirements - **Dense resource allocation:** Flex-start allocates GPU nodes physically close to each other within Google Cloud data centers, optimizing **dynamic machine learning network fabric** throughput and minimizing **inter-node latency**. - **Non-preemptible execution:** Unlike Spot VMs, once **Compute Engine** provisions a Flex-start instance group or cluster, the VMs run uninterrupted until the workload completes or reaches the maximum run duration of seven days. - **Flexible scheduling:** Flex-start queues provisioning requests and automatically instantiates the cluster as soon as the requested contiguous capacity becomes available. - **Cost efficiency:** It offers discounted pricing of **up to 53%** for vCPUs, memory, and accelerator types (such as **A3 Ultra** and **A4** GPUs) on a **pay-as-you-go (PAYG)** basis, avoiding charges for idle reservation time. ### Key advantages and capabilities - **Up to seven-day lifespan:** Perfectly matches multi-day fine-tuning runs and batch inference pipelines. - **Broad orchestration integration:** Can be provisioned directly via **Google Kubernetes Engine (GKE)**, **Slurm clusters**, **Cluster Toolkit**, or Compute Engine **Managed Instance Groups (MIGs)** with resize requests. - **Higher GPU obtainability:** Dynamic queuing gives workloads a higher probability of acquiring scarce accelerator hardware compared to standard on-demand requests. - **Preemptible quota utilization:** Leverages preemptible accelerator quotas rather than consuming expensive standard on-demand quotas. ### Why this is optimal Flex-start is specifically designed for non-fault-tolerant, short-duration training and fine-tuning jobs that have flexible start windows. It provides the low-latency network topology of reservations and significant cost savings without the ongoing financial overhead of dedicated calendar reservations.

Ai Hypercomputer Consumption Model Dynamic Scheduling | Intrigued by the art of cloud architecture? Discover how to design, develop, and manage robust, secure, scalable, and dynamic solutions on Google Cloud as you prepare for the Professional Cloud Architect exam!

IntermediateProfessional Cloud Architect

A machine learning engineering team is training a large-scale deep learning model on Google Cloud TPU accelerators using TensorFlow. During performance profiling with TensorBoard, the team observes that the TPU cores experience frequent idle periods and low Matrix Multiply Unit (MXU) utilization. Further analysis indicates that the TPU accelerator is waiting on frequent communication handshakes with the host CPU after every individual batch step. Which configuration change should the team make to reduce host-device communication overhead and maximize training throughput?

Using AI Hypercomputer (e.g., using AI Hypercomputer, Cloud Run functions, and Vertex AI for ML/AI workloads; integrating GPUs and TPUs in ML model training and serving; optimizing for different consumption models; and running large-scale AI model training

IntermediateProfessional Cloud Architect

An organization is designing an automated, event-driven machine learning workflow on Google Cloud. The architecture must satisfy the following technical requirements: Event-Driven Trigger: Automatically initiate the pipeline whenever a new training dataset is uploaded to a designated Cloud Storage bucket. Serverless Orchestration & Training: Run custom distributed model training scripts in custom containers using high-demand NVIDIA A100/H100 GPU accelerators via Vertex AI. Capacity Management & Cost Optimization: Accommodate GPU capacity constraints gracefully by queuing jobs and utilizing Dynamic Workload Scheduler (DWS) with a Flex Start scheduling strategy rather than failing immediately when on-demand quota is constrained. Automated Artifact Management: Upload model artifacts and deploy the trained model to a Vertex AI Endpoint. Which architectural approach should you implement?

IntermediateProfessional Cloud Architect

An enterprise is designing a centralized Customer-Managed Encryption Key (CMEK) strategy in Google Cloud to protect sensitive datasets across several workload projects. The security architecture must adhere to the following operational and compliance requirements: Separation of duties: Key administrators who create, configure automated rotation for, and retire cryptographic keys must not have permissions to perform cryptographic operations (encrypt/decrypt). Least privilege: Workload service accounts must only perform encryption and decryption operations on their specific designated keys without gaining access to other keys or administrative controls. Data continuity: Key rotation must occur automatically every 90 days without disrupting access to existing historical data encrypted with earlier key versions. Which architectural approach should you implement in Cloud Key Management Service (Cloud KMS)?

IntermediateProfessional Cloud Architect

An organization runs an image processing pipeline on Google Cloud Compute Engine. The system requires a guaranteed baseline capacity of 10 virtual machines running 24/7 to handle critical operational SLAs, while demand frequently spikes up to 40 additional virtual machines for fault-tolerant, stateless batch jobs that can tolerate interruptions and restarts. Which provisioning strategy offers the most cost-effective architecture while guaranteeing baseline performance SLAs?

IntermediateProfessional Cloud Architect

An enterprise is designing a modernization strategy for its distributed microservices architecture. The company currently operates services across on-premises data centers, AWS, and Google Cloud, but faces significant operational friction due to disparate deployment tooling, inconsistent security policies, and potential vendor lock-in. The architecture team has established the following technical requirements: Establish a common runtime layer that abstracts underlying infrastructure differences across environments Enable centralized, declarative configuration and security policy management from a single Git repository source of truth Manage multi-cluster deployments across on-premises and cloud environments as unified administrative units Which solution should you recommend to meet these requirements?

AdvancedProfessional Cloud Security Engineer

An enterprise organization implements a hub-and-spoke network architecture in Google Cloud. The architecture includes a centralized hub-vpc connected to an on-premises data center using Cloud Interconnect, and a separate spoke-vpc hosting application workloads. The two VPCs are interconnected using VPC Network Peering. The security and network engineering teams have established the following requirements: On-premises servers must be able to route traffic to the VM instances residing in spoke-vpc through hub-vpc. Cloud Router in hub-vpc must advertise the spoke-vpc address spaces to the on-premises Border Gateway Protocol (BGP) peers. Ingress firewall filtering in hub-vpc must restrict incoming traffic based on the identity of workloads running in spoke-vpc without using IP CIDR blocks or non-functional cross-peering constructs. Which combination of routing configurations and firewall controls should the organization implement?

Professional Cloud Architect

Ai Hypercomputer Consumption Model Dynamic Scheduling

Related questions