GPU VPS Offshore: Suc Manh Tinh Toan Iceland

AWS p3.2xlarge voi 1 GPU V100 co gia $3.06/gio = $2,203/thang neu chay lien tuc. Google Cloud va Azure cung tuong tu. Voi GPU VPS offshore tai Iceland, ban tra phi co dinh thap hon dang ke ma khong bi tinh tien moi gio. Hon nua, jurisdiction Iceland bao ve quyen rieng tu cua nghien cuu va du an AI nhay cam khoi yeu cau cung cap du lieu xuyen bien gioi.

Bon dang nguoi dung chinh huong loi:

• Nha nghien cuu doc lap: Training model duoi 7 ngay khong bi gian doan, khong bi cat khi tai khoan cloud bi flag. Iceland co the chay 24/7.
• Startup AI: Inference workload on dinh voi chi phi du doan duoc, khong tang vot khi traffic doi pho voi spike.
• Crypto mining cao cap: ETH Classic, Ergo, Ravencoin van profitable voi GPU NVIDIA. AnubizHost khong cam khoi luong mining trong gioi han ToS.
• Render studio: Blender, Unreal Engine 5 render farm voi chi phi diem san xuat thap hon nhieu so voi cloud lon.

Nang luong xanh cua Iceland (gan 100% tu thuy dien va dia nhiet) tao ra carbon footprint cuc thap - mot loi the cho cac du an AI quan tam den sustainability.

Sau khi nhan thong tin SSH, kiem tra GPU duoc gan dung:

lspci | grep -i nvidia
# 01:00.0 VGA compatible controller: NVIDIA Corporation ...

Cai dat NVIDIA driver va CUDA toolkit phien ban moi nhat tren Ubuntu 22.04:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
dpkg -i cuda-keyring_1.1-1_all.deb
apt update
apt install -y cuda-toolkit-12-3 nvidia-driver-535
reboot

Sau khi reboot, xac minh:

nvidia-smi
# +---------------------------------------------------------------------+
# | NVIDIA-SMI 535.xx     Driver Version: 535.xx     CUDA Version: 12.3 |
# +---------------------------------------------------------------------+

Cai dat Docker voi NVIDIA Container Toolkit de chay container co GPU access:

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg
curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \
  sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
  tee /etc/apt/sources.list.d/nvidia-container-toolkit.list
apt update && apt install -y nvidia-container-toolkit
nvidia-ctk runtime configure --runtime=docker
systemctl restart docker

Kiem tra: docker run --rm --gpus all nvidia/cuda:12.3.0-base-ubuntu22.04 nvidia-smi. Neu thay output cua nvidia-smi tu trong container, moi truong da san sang cho workload AI.

Cai PyTorch voi CUDA support trong conda environment:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b -p /opt/conda
source /opt/conda/etc/profile.d/conda.sh
conda create -n ai python=3.11 -y
conda activate ai
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install tensorflow[and-cuda] transformers accelerate

Jupyter Lab nen chay sau SSH tunnel hoac qua reverse proxy co authentication. KHONG mo Jupyter ra Internet truc tiep voi token mac dinh - day la nguon ro ri model va du lieu nghiem trong:

jupyter lab --generate-config
# Sua ~/.jupyter/jupyter_lab_config.py
c.ServerApp.password = u'sha1:hash_cua_password'
c.ServerApp.ip = '127.0.0.1'
c.ServerApp.open_browser = False
# Truy cap qua SSH tunnel: ssh -L 8888:localhost:8888 root@vps-ip

De train model lon dai ngay, dung tmux hoac screen de session tiep tuc chay khi mat ket noi SSH. Backup model checkpoint sang storage offsite qua rclone moi gio - mat 1 ngay training do crash con dau hon mat mot tuan luong.

Voi cau hinh nay, ban co moi truong AI/ML production tren GPU VPS Iceland kin dao, hieu nang cao va kinh te. AnubizHost ho tro mo rong tu 1 GPU len multi-GPU node khi du an phat trien.