Plan and manage an Azure AI solution 1.3 Manage, monitor, and secure AI systems
Your subscription has 240,000 TPM of quota for a model in East US.
Study guide skill: Manage quotas, scaling, rate limits, and cost footprints for model and agent workloads
Question B1-004
Your subscription has 240,000 TPM of quota for a model in East US. An existing deployment of that model in East US is assigned 200,000 TPM. You must create a second deployment of the same model in East US with 100,000 TPM. Which two actions could make this possible? (Choose two.)
Select 2 answers.
Answer: A, B
Quota is assigned per subscription, per region, per model and per deployment type in TPM, and each deployment takes TPM from that pool. Freeing TPM from the existing deployment or getting more quota both work. A new resource in the same subscription and region draws on the same quota, and max_tokens or a reader role doesn't change the quota.
Deep dive
How it works
Azure OpenAI quota is a pool, not a per-deployment allowance. Quota is assigned to your subscription on a per-region, per-model, per-deployment-type basis, measured in tokens per minute (TPM). Each time you create a deployment you assign it some TPM, and the available quota for that model drops by that amount. When the pool is used up, the docs give exactly two ways to create another deployment of that model: reduce the TPM assigned to other deployments of the same model, or request (and be approved for) a quota increase in that region.
The docs even use the same numbers as the question. With 240,000 TPM of quota in East US, you can create one 240K deployment, two 120K deployments, or any number of deployments in one or multiple resources, as long as the total stays under 240K in that region. Here 200K is already assigned, so only 40K is free. Lowering the existing deployment to 140K or less frees the 100K you need, and so does a quota increase.
The assigned TPM also becomes that deployment's rate limit, and a requests-per-minute (RPM) limit is set proportionally to it. So when you lower the existing deployment, it also gets less throughput.
The other options, and when you'd use them
- Lower max_tokens in the application's requests: max_tokens affects how the rate limiter estimates each request, not how much quota a deployment holds. The docs say to set it to the minimum value that serves your scenario to reduce 429s, but it doesn't free any quota.
- Create the deployment in a new Foundry resource in the same subscription and region: Quota is per subscription and region, not per resource. The docs say the TPM of deployments across one or multiple resources must add up within that region's quota, so a new resource draws from the same pool.
- Assign the Cognitive Services Usages Reader role: This role gives the minimal access needed to view quota usage across a subscription. It lets people see the numbers, but it doesn't change them.
Exam tip
Think "subscription + region + model + deployment type = one TPM pool." To fit a new deployment, either shrink another deployment of the same model or ask for more quota. New resources don't help.
References
Last reviewed: