CUDA Development Environment (Lambda Labs Native)

Simple CUDA development setup for Lambda Labs instances. No Docker required—Lambda provides CUDA pre-installed.

Quick Start

# Clone to persistent storage
cd /home/ubuntu/userspace
git clone https://github.com/5arast1/lambda-dev-env.git
cd lambda-dev-env

# Setup (one time)
chmod +x *.sh
./setup.sh

# Open new terminal or reload bashrc
source ~/.bashrc

# Validate
./validate.sh

Daily Usage

cd /home/ubuntu/userspace/lambda-dev-env/workspace

# Write your CUDA code
vim kernel.cu

# Compile (GH200/H100 = sm_90, A100 = sm_80)
nvcc -O3 -arch=sm_90 -o kernel kernel.cu

# Run
./kernel

# Profile
ncu --set full ./kernel

File Structure

lambda-dev-env/
├── setup.sh        # First-time setup
├── validate.sh     # Test environment
├── workspace/      # Your code (persisted)
└── scripts/        # Examples
    └── vector_add.cu

Compilation Cheatsheet

# Auto-detect GPU architecture
ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | tr -d ".")
nvcc -O3 -arch=sm_${ARCH} -o out in.cu

# Common architectures
nvcc -arch=sm_90 ...  # GH200, H100 (Hopper)
nvcc -arch=sm_89 ...  # L40 (Ada)
nvcc -arch=sm_86 ...  # A10 (Ampere)
nvcc -arch=sm_80 ...  # A100 (Ampere)

# With debug symbols
nvcc -G -g -arch=sm_90 -o out_debug in.cu

# Generate PTX
nvcc -ptx -arch=sm_90 in.cu

Profiling

# Nsight Compute (kernel profiling)
ncu ./kernel
ncu --set full ./kernel
ncu -o report --set full ./kernel  # Save report

# Nsight Systems (timeline)
nsys profile ./kernel
nsys profile -o timeline ./kernel

Lambda Labs Notes

• Persistent storage: /home/ubuntu/userspace survives instance stops
• CUDA version: 12.8 (pre-installed)
• GPU: GH200 480GB (sm_90, 132 SMs, 94.5GB HBM3)

Why No Docker?

Lambda's NFS-backed persistent storage doesn't support Docker's overlayfs operations. Since Lambda already provides a complete CUDA environment, Docker adds complexity without benefit. Use Docker only when you need reproducibility across different machines.

Example

cd workspace
cp ../scripts/vector_add.cu .
nvcc -O3 -arch=sm_90 -o vector_add vector_add.cu
./vector_add