Open Menu
Raven AI Engine v0.4.7: From an AI inference engine to a heterogeneous GPU computing platform

Raven AI Engine v0.4.7: From an AI inference engine to a heterogeneous GPU computing platform

User Name

Written by

Aleix Figueres

September 3, 2026

Table of Contents

Powering Fluendo AI Plugins v1.2.6 and production-ready Ubuntu GPU acceleration

During the past year, one of our main objectives at Fluendo has been evolving Raven AI Engine from what originally started as a hardware-accelerated AI inference engine into something significantly more ambitious: a heterogeneous GPU computing platform capable of orchestrating rendering, AI inference, and image processing as a single GPU-driven execution pipeline.

That architectural evolution reaches an important milestone with Raven AI Engine v0.4.7, which becomes the technological foundation of Fluendo AI Plugins (FAIP) v1.2.6. In addition to introducing our first production-ready Ubuntu GPU release, this version fundamentally changes how multimedia AI pipelines are built and maintained through programmable rendering, unified shader development, and a heterogeneous execution model designed around devices rather than APIs.

Motivation

When Raven AI Engine launched, its primary mission was straightforward: execute neural network inference efficiently across different hardware accelerators.

As the engine matured and became the execution core of Fluendo AI Plugins, we realized that inference was only one stage of a much larger multimedia pipeline. A production video application performs decoding, color conversion, preprocessing, inference, post-processing, rendering, synchronization, and optimization; only the neural network leaves a considerable amount of performance untapped.

Our objective, therefore, shifted from accelerating inference to keeping the entire multimedia pipeline on the GPU whenever possible.

This release represents the first production-ready implementation of that vision.

Rather than considering rendering and AI as two independent worlds, Raven now orchestrates them as a single heterogeneous workload capable of sharing resources, synchronization primitives, and execution scheduling across different compute and graphics technologies.

A Fully GPU-Driven Architecture

One of the biggest conceptual changes introduced during the Raven AI Engine 0.4.x series is that the engine is no longer organized around graphics APIs or AI frameworks.

Instead, Raven is organized around devices.

Render devices, AI devices, and CPU devices expose capabilities rather than technologies, allowing Raven to schedule work where it makes the most sense while keeping a consistent programming model across Windows and Linux.

The resulting execution model looks conceptually like this:

                    Raven Scheduler
                           │
          ┌────────────────┴────────────────┐
          │                                 │
   Render Devices                     AI Devices
 ───────────────────             ──────────────────
 Vulkan                         CUDA
 DirectX 12                     DirectML
 CPU / SwiftShader              CPU

This architecture allows the Raven AI Engine to compose heterogeneous pipelines in which rendering, preprocessing, inference, and post-processing remain coordinated without exposing backend-specific complexity to the upper layers.

On Linux, Vulkan becomes the owner of rendering resources while CUDA focuses on inference. Rather than repeatedly transferring textures through system memory, Raven AI Engine now keeps processing on GPU resources whenever possible, significantly simplifying the execution model while reducing unnecessary synchronization and data movement.

This architectural shift is considerably more important than simply “adding Vulkan support”. It establishes the foundation for future heterogeneous workloads in which different hardware accelerators collaborate within the same execution graph.

Slang and the birth of user programmable RSL Shaders

Perhaps the most exciting change introduced by Raven AI Engine v0.4.7 is not immediately apparent to end users.

Historically, implementing a new rendering effect meant modifying backend-specific graphics code. Supporting DirectX 12, Vulkan, CUDA, and CPU implementations quickly resulted in duplicated logic that became increasingly difficult to maintain.

Raven now solves this problem by combining Slang with its own RSL (Raven Shader Language) framework.

Slang provides backend portability while RSL introduces a controlled shader layer that allows rendering algorithms to be authored once and automatically compiled for the supported execution backends.

Even more importantly, those shaders are no longer internal implementation details.

They become production assets.

Every effect lives as an RSL source file that can be opened directly inside the Raven Shader Editor, modified interactively, validated visually, and automatically embedded into the final shared library during compilation.

Developer
     │
     ▼
  *.rsl shader
     │
     ▼
 Raven Shader Editor
     │
     ▼
 Automatic compilation
     │
     ▼
 Embedded into DLL / SO
     │
     ▼
 Production execution

This effectively creates a What You Edit Is What You Ship workflow.

The exact shader that engineers visualize during development is the same shader executed inside the deployed plugin.

Although this release introduces new anonymization effects built on top of this infrastructure—including higher-quality Gaussian blur, enhanced pixelation, blackout mode, configurable fade-out regions, and overlay rendering—the real innovation is architectural rather than visual.

From this point onwards, many rendering capabilities can evolve simply by developing new user shaders instead of introducing new rendering code throughout the engine.

In practice, RSL becomes the foundation for future GPU effects, overlays, visual debugging tools, and customer-specific rendering extensions while dramatically reducing maintenance costs across multiple graphics backends.

What this means for Fluendo AI Plugins v1.2.6

The architectural improvements introduced in Raven immediately benefit Fluendo AI Plugins.

Version 1.2.6 becomes our first production-ready Ubuntu release powered by Raven’s heterogeneous GPU execution model. Rendering, preprocessing, AI inference, and image processing are now orchestrated by the same execution engine, providing a significantly cleaner architecture than previous releases.

The Face Anonymizer has also been migrated to the new programmable rendering framework, introducing higher-quality Gaussian blur, improved pixelation, blackout mode, configurable anonymization enlargement, configurable edge fade-out, overlay rendering capabilities, and new detection threshold controls. Because these effects are implemented as RSL shaders, they are considerably easier to evolve, validate, and maintain.

This release also improves deployment by redistributing the CUDA runtime and cuDNN libraries together with the application, leaving the NVIDIA Driver as the only external runtime prerequisite. Combined with dedicated runtime validation environments, expanded IT-Suite automated regression testing, and complete plugin documentation distributed with every bundle, the deployment experience becomes significantly more predictable across supported platforms.

From a production engineering perspective, perhaps the most important achievement is not a single feature but the platform that now exists underneath them. Raven is no longer simply executing AI models—it provides a programmable heterogeneous GPU platform on which future multimedia AI capabilities can be developed considerably faster.

Current limitations and future work

Although Raven IA Engine’s Vulkan backend has now reached production quality, the surrounding GStreamer Vulkan ecosystem still presents maturity limitations, particularly regarding decoder integration across different drivers and distributions.

For this reason, NVIDIA hardware decoders remain the recommended execution path for Ubuntu deployments in Fluendo AI Plugins v1.2.6.

Support for VA-API hardware decoding is planned for the next major release (v1.2.5), together with continued work on expanding heterogeneous scheduling, GPU interoperability, and programmable rendering capabilities through RSL.

The introduction of user-programmable shaders also opens the door to significantly richer rendering pipelines, allowing future effects to be implemented almost entirely at the shader level without requiring invasive engine modifications.

While Ubuntu GPU acceleration is the most visible outcome of this release, the real milestone is the underlying architectural transformation that makes those future developments considerably easier.

Experience the future of GPU-Driven multimedia

Raven AI Engine v0.4.7 and Fluendo AI Plugins v1.2.6 are now ready to power your production workloads on Ubuntu. If you want to eliminate CPU-GPU bottlenecks and accelerate your AI video pipelines:

Download and try Fluendo AI Plugins 🚀

For further information, get in touch with our team to discuss how we can help optimize your custom GStreamer and GPU pipelines.