Generative AI Leader Practice Question: A research team is training a large multimodal…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Compute Engine with A100 GPUs

  Why it's wrong here

  Compute Engine with A100 GPUs offers general-purpose accelerated VMs, but the fabric and scheduling are not purpose-built for the tightly coupled, multi-host training a large multimodal model requires. It suits flexible GPU workloads. Dedicated training infrastructure delivers the interconnect bandwidth and scale this scenario needs.

• ✓

  TPU Pods

  Why this is correct

  TPU Pods interconnect thousands of Tensor Processing Units over dedicated high-bandwidth optical links, enabling the massive parallel matrix operations that large multimodal training demands. This specialised architecture delivers far higher throughput for a fixed budget than general-purpose GPU clusters, directly minimising the training time constraint stated in the scenario.

• ✗

  Kubernetes Engine with GPU nodes

  Why it's wrong here

  Kubernetes Engine with GPU nodes orchestrates general containerised workloads; it lacks the purpose-built interconnects and compiler stack for large-scale distributed training. It suits serving or managing mixed workloads. Minimising training time for a large multimodal model needs Google's dedicated training infrastructure.

• ✗

  Cloud TPU v5e single chip

  Why it's wrong here

  A single v5e chip provides one accelerator, insufficient for the distributed, multi-host training a large multimodal model demands within a fixed budget. It suits inference or fine-tuning smaller models. Large-scale training requires a TPU pod or slice interconnecting many chips via high-speed interconnects.

Go deeper

Related to this question