---
title: "GPU Node FAQ"
description: "Quick answers about GPU availability, cards, snapshots, and billing, each linking to the page with the full detail."
content_type: "faq"
product: "gpu-node"
platform: "MyAccount"
url: "/myaccount/gpu-node/faq/"
---

# GPU Node FAQ

## Should I use TIR instead of a GPU node for AI and machine learning work?

Yes, for most AI and machine learning work. [TIR](/tir/) is E2E's platform for it. A TIR Instance
starts with a pre-built image for PyTorch, TensorFlow, Transformers, vLLM, and more, with JupyterLab
ready. TIR offers GPUs in Delhi and Chennai, and Hopper and later GPUs, the H100, H200 and B200, are on TIR only. A GPU node
suits a GPU that has to sit beside your other MyAccount resources, or a Windows GPU machine. See
[Get started with TIR Instances](/tir/tir-instance/get-started/).

## Is a GPU node available in Chennai?

Not as a MyAccount GPU node: those are offered in Delhi, so switch your region to Delhi to see
the **GPU** tab. For a GPU in Chennai, use [TIR](/tir/), which runs GPU instances there. See
[GPU Node](/myaccount/gpu-node/).

## Which GPU card should I pick?

It depends on your workload and model size: an A100 80GB for large-memory LLM
inference and training, an L40S for cost-efficient inference, an A40 for visualization and
rendering, an A30 or L4 for smaller-model or lower-power inference, or a V100 or T4 to match an
existing pipeline. See
[Choose a GPU card](/myaccount/gpu-node/concepts/choose-a-gpu-card/) for the full comparison.

## Can I snapshot a GPU node?

Yes, on **GDC3** nodes, which support snapshots the same as any node. On **GDC** nodes, the
other GPU family (see [Node families](/myaccount/node/concepts/node-families/)), use
**Save Image** instead to capture a re-launchable copy of the root disk. See
[Save and reuse a GPU image](/myaccount/gpu-node/how-to/save-and-reuse-a-gpu-image/).

## Can I attach a volume when I create a GPU node?

Usually not: most GPU plans hide the Volume step during creation. Create the node first, then attach the
volume from its details page. See
[Create a GPU node](/myaccount/gpu-node/how-to/create-a-gpu-node/).

## How is a GPU node billed?

By the minute, at its plan's hourly price, the same as any node, with no minimum charge. A
committed plan is paid for its whole term instead. See
[Committed nodes](/myaccount/node/concepts/committed-nodes/).

## Can I upgrade a GPU node to a different card?

No. You can upgrade only within the same card, for example a 1x plan to a 4x plan on the same card.
To move to a different card model, save an image and launch a new node on the target card. See [Save
and reuse a GPU image](/myaccount/gpu-node/how-to/save-and-reuse-a-gpu-image/).

## Do I need to manage the NVIDIA driver myself?

No. The driver ships with the node's image. `nvidia-smi`, run right after connecting, confirms
it is working. If it is not, see
[Troubleshooting GPU Node](/myaccount/gpu-node/troubleshooting/gpu-node/).

## What is the best way to run a GPU workload?

In a container: the host provides the driver, and the container carries CUDA and your
framework. See [Run GPU workloads in Docker](/myaccount/gpu-node/how-to/run-gpu-workloads-in-docker/).

## How do I serve an LLM from a GPU node?

Run vLLM or Hugging Face TGI in a container, behind an OpenAI-compatible API. See [Serve LLM
inference on a GPU node](/myaccount/gpu-node/how-to/serve-llm-inference-on-a-gpu-node/). For managed
model endpoints, see [TIR](/tir/).