Skip to content

Hardware Acceleration & Image Flavors

FileFlows Real Image provides specialized, vendor-optimized container image flavors engineered for high-performance GPU transcoding. Each flavor pre-bakes the vendor-specific runtime and driver stack at build time, eliminating runtime package installations (apt-get) and stripping out unneeded driver bloat from competing vendors.

Image Flavors Overview

Flavor Tag Target Hardware / Architecture Content Size Virtual Size Driver Stack Included
:intel Intel Core Gen 8–14+, Arc Alchemist, Battlemage, N-series 514 MB 1.76 GB Intel Media Driver (iHD 26.1+), Level Zero (libze), oneVPL, OpenCL ICD
:amd AMD Radeon RX 5000–8000 series, Ryzen 6000–9000 APUs 473 MB 1.65 GB Mesa Gallium (radeonsi), RADV Vulkan, AMDGPU DRM
:cuda NVIDIA Pascal through Ada Lovelace (Host-Based CUDA) 473 MB 1.65 GB Host-injected driver hooks (libcuda, NVENC, NVDEC) with zero package bloat
:cuda13 NVIDIA Ada Lovelace, Blackwell (RTX 50xx), Hopper (CUDA 13.4) 720 MB 2.33 GB Minimal NVIDIA CUDA 13.4 runtime + NVRTC & NPP video filters
:latest / :all Universal multi-vendor default (Intel + AMD + NVIDIA runtimes) 574 MB 2.00 GB Full Intel Media Driver, Mesa Gallium VA-API, and NVIDIA host driver hooks

1. Intel QuickSync & Arc GPUs (:intel)

Supported Hardware

  • Integrated Graphics: Intel Core 8th Gen (Coffee Lake) through 14th Gen (Raptor Lake Refresh), Intel Core Ultra (Meteor Lake, Arrow Lake, Lunar Lake), and Intel N-series (N100, N200, N305).
  • Discrete GPUs: Intel Arc Alchemist (A310, A380, A750, A770) and Intel Arc Battlemage (B570, B580).

Pre-Baked Driver Stack

The :intel flavor is built specifically for Intel hardware with zero AMD or NVIDIA dependencies:

  • intel-media-va-driver-non-free (v26.1.2+): Intel Media Driver (iHD) with native dual support for both i915 and modern Linux xe kernel DRM drivers (xe.ko). Accelerates AV1, HEVC 10-bit, and H.264 encode/decode.
  • libze-intel-gpu1: Intel oneAPI Level Zero GPU driver for modern Xe and Xe2 architectures.
  • intel-opencl-icd: Intel Compute Runtime for hardware-accelerated OpenCL HDR tone mapping (HDR10 to SDR) and color grading filters.
  • libvpl2 & libmfx-gen1.2: Intel oneVPL runtime for direct QSV transcode pipelines.
  • Zero Legacy Baggage: Deprecated drivers (such as i965 for pre-2015 CPUs) are purged.

Compose Configuration (:intel)

services:
  fileflows:
    image: ghcr.io/lusoris/fileflows-real-image:intel
    container_name: fileflows
    restart: unless-stopped
    devices:
      - /dev/dri:/dev/dri
    environment:
      - PUID=1000
      - PGID=1000
      - TZ=UTC

2. AMD Radeon & Ryzen APUs (:amd)

Supported Hardware

  • Radeon Discrete GPUs: RDNA 1 (RX 5000), RDNA 2 (RX 6000), RDNA 3 / 3.5 (RX 7000), and RDNA 4 (RX 8000+).
  • Ryzen APUs: Ryzen 4000/5000 (Vega), Ryzen 6000/7000/8000/AI 300 APUs with dual hardware AV1 engines.

Pre-Baked Driver Stack

The :amd flavor slashes image content size down to 473 MB (-53% vs upstream):

  • mesa-libgallium: Official Mesa 26.0+ Gallium driver providing /usr/lib/x86_64-linux-gnu/dri/radeonsi_drv_video.so for hardware-accelerated VA-API H.264, HEVC, and AV1 encode/decode.
  • mesa-vulkan-drivers (RADV): High-performance AMD Vulkan driver for Vulkan-based video filters and compute.
  • libdrm-amdgpu1: Direct AMDGPU kernel DRM interface.
  • Zero Intel/NVIDIA Bloat: Omits all Intel Media Drivers and NVIDIA CUDA toolkits.

Compose Configuration (:amd)

services:
  fileflows:
    image: ghcr.io/lusoris/fileflows-real-image:amd
    container_name: fileflows
    restart: unless-stopped
    devices:
      - /dev/dri:/dev/dri
    environment:
      - PUID=1000
      - PGID=1000
      - TZ=UTC

3. Host-Based NVIDIA CUDA (:cuda)

Supported Hardware

  • NVIDIA Pascal (GTX 10xx), Turing (GTX 1660, RTX 20xx), Ampere (RTX 30xx), and Ada Lovelace (RTX 40xx).

Architecture & Driver Stack

The :cuda flavor delivers maximum efficiency by stripping out all in-container package bloat (such as unneeded cuBLAS and cuFFT AI libraries) and relying directly on host driver injection:

  • Uses the official NVIDIA Container Toolkit to dynamically bind-mount libcuda.so.1, libnvidia-encode.so.1 (NVENC), and libnvcuvid.so.1 (NVDEC) at container startup.
  • Configured with NVIDIA_DRIVER_CAPABILITIES=compute,video,utility and NVIDIA_VISIBLE_DEVICES=all.
  • Drops image content size down to 473 MB and virtual disk size to 1.65 GB (slashing over 4.3 GB of wasted bloat).

Compose Configuration (:cuda)

services:
  fileflows:
    image: ghcr.io/lusoris/fileflows-real-image:cuda
    container_name: fileflows
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - PUID=1000
      - PGID=1000
      - TZ=UTC

4. Minimal Cutting-Edge NVIDIA CUDA 13.4 (:cuda13)

Supported Hardware

  • Ada Lovelace (RTX 4070, 4080, 4090 with dual 8th-Gen NVENC AV1 encoders).
  • Blackwell (RTX 5070, 5080, 5090, B100, B200 with 9th-Gen NVENC).
  • Hopper (H100, H200) and Ampere (RTX 30xx).

Why CUDA 13.4 Minimal?

CUDA 13.4 introduces optimized compute kernels and modern hardware acceleration pipelines tailored for next-generation architectures like Blackwell (RTX 50xx). The :cuda13 image is engineered strictly with the video transcode runtime essentials:

  • cuda-nvrtc-13-4: NVIDIA Runtime Compilation library required for dynamic FFmpeg CUDA filter compilation (scale_cuda, yadif_cuda, overlay_cuda, etc.).
  • cuda-cudart-13-4: Core CUDA 13.4 runtime API library.
  • libnpp-13-4: NVIDIA Performance Primitives for NPP-based video filtering and color processing (scale_npp).
  • Zero AI Math Bloat: Excludes libcublas (-850 MB), libcusolver (-390 MB), libcusparse (-380 MB), libcufft (-273 MB), and cuda-compat (-200 MB), slashing over 2.7 GB of unneeded libraries while retaining 100% video transcoding filter support.

Compose Configuration (:cuda13)

services:
  fileflows:
    image: ghcr.io/lusoris/fileflows-real-image:cuda13
    container_name: fileflows
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - PUID=1000
      - PGID=1000
      - TZ=UTC

5. Universal Default Image (:latest)

If you operate a mixed cluster with both Intel and AMD nodes, or prefer a single image that adapts to whatever GPU is available, use ghcr.io/lusoris/fileflows-real-image:latest. It pre-bakes the full Intel QuickSync stack, Mesa Gallium VA-API for AMD, and environment hooks for NVIDIA Container Toolkit, all within a compact 574 MB content footprint.

Verification & Diagnostics

VA-API (Intel / AMD)

Verify that your GPU encoder is accessible inside the container by running vainfo:

docker exec -it fileflows vainfo

Sample output confirming AV1, HEVC, and H.264 hardware encoding:

VAProfileH264Main    : VAEntrypointEncSlice
VAProfileHEVCMain    : VAEntrypointEncSlice
VAProfileAV1Profile0 : VAEntrypointEncSlice

NVIDIA (NVENC / NVDEC)

Verify that your NVIDIA GPU is recognized by running nvidia-smi:

docker exec -it fileflows nvidia-smi

Troubleshooting /dev/dri/renderD128 Permissions

If FileFlows cannot access the GPU device:

  1. Verify device permissions on the host:
ls -la /dev/dri
  1. Identify the host group ID for render (typically 107 or 993 on Ubuntu/Debian).
  2. If necessary, add the host group ID via group_add in docker-compose.yml:
services:
  fileflows:
    group_add:
      - "993"