To configure the performance of an Azure OpenAI model deployment using an ARM template, you must set the capacity property within the sku object. This property defines the model's throughput in thousands of Tokens-Per-Minute (kTPM).

Azure OpenAI Service enforces a rate limit ratio of 6 Requests-Per-Minute (RPM) for every 1,000 Tokens-Per-Minute (TPM). To meet the requirement of 600 RPM, you must calculate the corresponding TPM:

1. Calculate required TPM: Required TPM=6Required RPM​×1,000 Required TPM=6600​×1,000=100×1,000=100,000 TPM
2. Convert TPM to the capacity value: The capacity property is measured in units of 1,000 TPM. capacity=1,000100,000 TPM​=100

Therefore, setting the capacity to 100 correctly provisions the resource to handle 600 requests per minute.

Microsoft Learn, Azure OpenAI Service quotas and limits. This document details the relationship between Tokens-Per-Minute (TPM) and Requests-Per-Minute (RPM). In the section "Managing quota," it states, "For standard deployments the requests-per-minute (RPM) limit is proportional to the tokens-per-minute (TPM) limit assigned... the ratio between TPM and RPM is 6 RPM per 1,000 TPM."

Microsoft Docs, Bicep & ARM template reference for Microsoft.CognitiveServices/accounts/deployments. The official template reference specifies the valid properties for the sku object of a deployment. It lists name (string) and capacity (integer) as the key properties for defining the deployment's performance tier and throughput.