Quickstart: from checkout to your first GPU job

This guide takes you from a delivered server to a running GPU job. It assumes you have reserved a server and paid the first month; you received the hostname by email.

1. Add your SSH key

Servers accept SSH keys only, never passwords. Paste your public key during checkout, or send it to support before delivery. To create a key:

ssh-keygen -t ed25519 -C "you@company.com"
cat ~/.ssh/id_ed25519.pub

2. Connect

ssh root@your-server.coregpu.net

You are root on a dedicated machine. Create named accounts for your team with adduser and add their keys to ~/.ssh/authorized_keys.

3. Check the GPUs

On NVIDIA servers:

nvidia-smi                      # GPUs, memory, driver and CUDA version
nvidia-smi topo -m              # NVLink / PCIe topology between GPUs
nvidia-smi nvlink --status      # per-link NVLink state

On AMD Instinct servers:

rocm-smi                        # GPUs, memory and utilisation
rocm-smi --showtopo             # Infinity Fabric topology
amd-smi list                    # detailed device list

On an 8-GPU SXM or OAM server, every GPU pair should show an NVLink (NV) or XGMI link in the topology matrix.

4. Run a first job

Our ready images include the drivers and PyTorch. To check that PyTorch sees every GPU:

python3 -c "import torch; print(torch.cuda.device_count(), torch.cuda.get_device_name(0))"

Prefer containers? Docker and the GPU container runtime are preinstalled:

# NVIDIA
docker run --rm --gpus all nvidia/cuda:13.0.0-base-ubuntu24.04 nvidia-smi

# AMD
docker run --rm --device=/dev/kfd --device=/dev/dri --group-add video \
  rocm/pytorch:latest python -c "import torch; print(torch.cuda.device_count())"

5. Benchmark the interconnect

Before a long run, measure collective bandwidth with NCCL tests (RCCL tests on AMD):

git clone https://github.com/NVIDIA/nccl-tests.git && cd nccl-tests
make -j
./build/all_reduce_perf -b 8 -e 8G -f 2 -g 8

Large-message bus bandwidth should approach the NVLink figures listed on the GPU pages. Our guide How to validate a new 8-GPU server covers burn-in and health checks in detail.

6. Store data and checkpoints

Every server has local NVMe scratch storage. It is fast and private to your server, but it is erased when the server is returned: copy checkpoints you want to keep to your own object storage before the end of your month.

7. Renew or stop

Servers renew monthly. You can renew, resize or stop at the end of any month; see Billing.