Plan and manage an Azure AI solution 1.2 Set up AI solutions in Foundry
A trading platform runs a mission-critical, high-volume workload and needs guaranteed…
Study guide skill: Choose appropriate deployment options
Question B1-003
A trading platform runs a mission-critical, high-volume workload and needs guaranteed throughput with consistent latency. Cost predictability matters more than paying only for tokens used. Which deployment option fits best?
Select one answer.
Answer: A
Provisioned throughput reserves capacity, billed per PTU per hour, and is intended for mission-critical, high-scale production workloads that need guaranteed throughput and consistent latency. Standard deployments suit variable traffic, and batch is discounted asynchronous processing with no low-latency target.
Deep dive
How it works
Provisioned throughput gives your deployment dedicated model processing capacity. With standard deployments, inference capacity is shared across customers and throughput can vary with demand. A provisioned deployment holds a fixed amount of capacity exclusively for you, whether or not requests are being made. It's billed per PTU per hour (or through Azure reservations), and the docs describe it as built for mission-critical, high-scale production workloads that need guaranteed throughput and consistent latency.
Capacity is measured in provisioned throughput units (PTUs). When you create a provisioned deployment, you choose how many PTUs to allocate, and Foundry reserves that compute for you. PTU quota isn't tied to one model: the same quota can be used to deploy any supported model. The docs say provisioned throughput fits predictable traffic, latency-sensitive requirements, production-scale volume where per-token billing becomes expensive, and real-time or interactive apps. One planning note: capacity changes throughout the day, so always verify capacity availability before you plan a deployment or buy a reservation.
The other options, and when you'd use them
- Standard (pay-per-token) deployment: You pay per token and there's no latency SLA. Standard remains the better fit for development, testing, low-volume usage, or traffic that's too variable to size in advance.
- Batch deployment: Batch is billed per token at a discounted rate, with no latency requirements, and results come back asynchronously. Use it for bulk processing that isn't time-sensitive.
- Developer deployment: This type is for fine-tuned model evaluation only, with no SLA.
Exam tip
"Guaranteed throughput", "consistent latency" and "predictable cost" point to provisioned throughput. "Variable or unpredictable traffic" points to standard, and "bulk jobs that can wait" points to batch.
References
Last reviewed: