NOVABERDocs

Loading search index…

    Documentation menu

    Getting started

    Novaber system, GPU and storage requirements

    Novaber · Content reviewed

    On this page

    Requirements depend on the selected app. The app requirements below come from the current public catalog; one app's VRAM requirement does not apply to all apps.

    NVIDIA driver and CUDA baseline

    Novaber NVIDIA GPU workflows require driver major version >= 575. For this project's CUDA 12.9.1 baseline, the Windows driver must also be 576.57 or newer. Read the complete driver version and compare numerically; a successful nvidia-smi command alone does not pass this check.

    nvidia-smi --query-gpu=name,driver_version,memory.total --format=csv,noheader,nounits
    

    For example, 574.99 fails the major-version requirement; 575.90 and 576.02 fall below the Windows baseline; 576.57 and 610.47 meet this driver policy. Unavailable or unparseable versions remain unknown; denied access remains restricted. Neither passes. Preserve satisfactory newer drivers, then verify container GPU access and the selected app separately.

    This is Novaber project policy, not a universal requirement for every CUDA 12.x application. NVIDIA lists corresponding CUDA 12.9 Update 1 drivers as Linux 575.57.08 and Windows 576.57; limited minor-version compatibility on older drivers does not override the project baseline. See the CUDA 12.9.1 release notes. A complete Windows CUDA Toolkit installation is not required for this container workflow.

    For an assisted check, follow AI setup with the documentation site’s setup Skill 1.0.3. Keep the references, diagnostic script and checksum manifest from that same skill version. Record desktop, app and skill versions separately. After the driver check, verify container GPU access and an actual task using the workflow.

    Desktop and runtime

    • Windows 10 / 11 x64: An x64 desktop installer is currently available.

    • Docker Desktop: Downloads images and creates and runs application containers.

    • NVIDIA GPU: GPU and video memory requirements vary by application.

    An internet connection is needed to download applications and sign in. Download size and setup time depend on the image and your connection.

    Hardware and storage for released apps

    AppGPUVRAMImages and models
    3D ReconstructionNVIDIA GPU Required12GB+ VRAMImage >10GB
    Cutout & Watermark RemovalNVIDIA GPU required8 GB+ VRAM recommendedImage downloads ~16 GB
    OCR Document RecognitionNVIDIA GPU required (CUDA 12.6)VRAM varies by model and page complexityBackend/worker ~18.1 GB unpacked each; models extra
    Voice StudioNVIDIA GPU recommended4 GB standard voices; 8 GB recommended for clone/designShared image layers use about 7 GB; the three models require about 9.6 GB more

    The desktop client targets Windows 10 / 11 x64. Docker Desktop, WSL and GPU drivers must also meet the selected application's requirements and current official support requirements.

    Current Docker Desktop Windows requirements ↗

    Interpret the requirements correctly

    VRAM is GPU memory; system RAM and disk space do not directly replace it. Required and recommended wording follows the app catalog. If the catalog gives no fixed VRAM minimum, this guide does not invent one.

    Download size, extracted image size and additional model size measure different things. They are not directly comparable and do not represent total disk use. Docker data, model caches, inputs and exports can live on different drives. Check the actual drives and allow room for extraction, temporary files and results. No universal total footprint or multiplier has been measured.

    Decide whether your computer is suitable

    1. Check Windows x64 and the current Docker Desktop support policy. Being able to install WSL does not prove the Windows release is still supported by Docker.
    2. Check the selected app’s GPU and VRAM requirements. A recommendation is not evidence of a tested CPU alternative.
    3. Verify WSL, Docker and container GPU access with the installation guide, then check a small task and export using the workflow.

    These requirements help screen equipment, without guaranteeing every workload size. Resolution, sequence length and model choice may change resource use; inspect actual results and errors.

    Inspect the existing environment: measured capacity or a passing host GPU check does not establish readiness for the selected app, model and input size.

    Sources