Docs
Documentation.
Everything you need on your server: first login, SSH, drivers, CUDA and ROCm, containers, multi-GPU jobs, serving and data.
9 guides · Pricing API reference
Get started
- Your first hour on a new GPU server Log in to your new GPU server, check its GPUs, CPU, memory and disks, update the OS and create a sudo user. Commands for Ubuntu 22.04 and 24.04 LTS. 4 min read · 6 sections
- SSH access: keys, firewall, tunnels and file transfer Set up ed25519 keys and ~/.ssh/config, turn off password logins, add ufw and fail2ban, keep jobs alive in tmux, tunnel Jupyter and copy data with rsync. 4 min read · 8 sections
- Billing, crypto payments and renewals How CryptGPU billing works: prepaid monthly terms in USD, paying an invoice in BTC, ETH, USDT, USDC or LTC, renewals, changing servers, end of term and refunds. 3 min read · 8 sections
GPU software
- NVIDIA drivers, CUDA and Fabric Manager on Ubuntu Install the NVIDIA driver and CUDA toolkit on Ubuntu, enable persistence mode, run Fabric Manager on HGX servers, check NVLink, use MIG and fix Xid errors. 7 min read · 7 sections
- ROCm on AMD Instinct MI355X: install and verify Install the AMD GPU driver and ROCm on an MI355X server, check it with rocminfo and amd-smi, run PyTorch and ROCm containers, choose GPUs and test RCCL. 5 min read · 8 sections
- Docker with GPUs: NVIDIA and ROCm containers Install Docker and the NVIDIA Container Toolkit, run NGC and ROCm images with GPU access, reserve GPUs in Docker Compose and keep data on local NVMe volumes. 5 min read · 7 sections
Workloads
- Multi-GPU on one server: topology, NCCL and torchrun Read the GPU topology, set useful NCCL variables, launch 2, 4 or 8 GPU jobs with torchrun, measure bus bandwidth with nccl-tests and use RCCL on AMD MI355X. 5 min read · 7 sections
- Serving LLMs with vLLM on dedicated GPUs Serve open models with vLLM on 1 to 8 GPUs: tensor parallelism, FP8 and 4-bit quantization, an OpenAI-compatible API behind SSH or TLS, and GPU monitoring. 5 min read · 7 sections
- Local NVMe, RAID 0 and moving data in and out Identify, format and mount NVMe drives by UUID, build an optional RAID 0 with mdadm, move datasets with rsync and rclone, and back up before your term ends. 4 min read · 7 sections
No guide matches. Try another word, or ask us.
Pricing API
Public JSON endpoints for our GPU catalog, prices and quotes. No key needed.
GET /api/v1/gpus
API reference
# log in
$ ssh root@<server-address>
# GPUs, memory and utilization
$ nvidia-smi # NVIDIA
$ rocm-smi # AMD
# GPU topology (NVLink, PCIe)
$ nvidia-smi topo -m
# a container that sees every GPU
$ docker run --rm --gpus all ubuntu nvidia-smiStuck on something? Ask us.
Tell us your GPU, what you ran and what you saw. We reply by email.