BrightDrive wants to replace a low-traffic real-time GPU endpoint with SageMaker Serverless Inference for cost savings. They need to control concurrency and understand latency trade-offs. Which statement most accurately describes how to configure serverless and the trade-offs compared with a persistent real-time endpoint?

The correct option accurately describes SageMaker Serverless Inference characteristics. You configure MaxConcurrency and memory size, and it scales from zero, leading to cold-start latency for initial requests after idle periods. This makes it cost-effective for intermittent, low-traffic workloads. In contrast, persistent real-time endpoints offer lower, more consistent latency, support GPUs, and allow for provisioned capacity, making them suitable for high-traffic or latency-sensitive applications. The incorrect options contain factual errors: Setting ProvisionedConcurrency on Serverless Inference helps mitigate cold starts but Serverless Inference does not support GPU instance types. Serverless Inference does have concurrency settings (MaxConcurrency) and can handle parallel invocations, not just serial workloads. Switching to Serverless does not guarantee identical p99 latency, especially due to cold starts, even with MaxConcurrency set. Cold starts can occur even if MaxConcurrency 1 if the endpoint scales down to zero.