Amazon EC2 Accelerated Instances
GPU and AI-accelerator instances for training and inference.
❓ What is it?
EC2 instance families with hardware accelerators: P-series (NVIDIA GPUs for training), G-series (graphics/inference GPUs), and Trn/Inf series built on AWS's own Trainium and Inferentia chips.
💡 Why does it exist?
Deep learning is matrix math at massive scale; CPUs are the wrong tool. Accelerators cut training from weeks to hours — and choosing the RIGHT family is the difference between a sensible bill and a shocking one.
⏱️ When should you use it?
Choose P/Trn instances for heavy training, G/Inf for cost-efficient inference; only when a managed layer (SageMaker, Bedrock) does not already abstract the hardware away.
🗺️ Where does it fit?
Underneath training jobs and self-managed model servers — usually indirectly, as the instance types you select inside SageMaker configuration.
🔌 How do you integrate it?
Select the instance type in SageMaker jobs/endpoints or launch EC2 with a Deep Learning AMI; reduce cost with Spot for interruptible training and Savings Plans for steady inference.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Family mnemonics: P = training GPUs, G = inference/graphics GPUs, Trn = Trainium (training), Inf = Inferentia (inference).
- Spot instances for fault-tolerant training is the recurring cost-optimisation answer.