Technical Intelligence Report
Date: 2025-12-10

🤖 ROCm Updates & Software

[2025-12-10] AMD FSR™ Redstone expands with the latest AMD FSR™ SDK 2.1

Source: AMD GPUOpen

Key takeaway relevant to AMD: > AMD is pivoting heavily toward ML-based “neural rendering” with the launch of FSR SDK 2.1, introducing a modular DLL-based architecture that allows for driver-level updates of upscaling and frame generation technologies without requiring game-specific patches.

Summary: > AMD has released FSR SDK 2.1, the official launchpad for “Redstone” technologies. This release transitions FSR from traditional analytical methods to ML-powered solutions including Upscaling, Frame Generation, Ray Regeneration, and Radiance Caching. It introduces a modular system using specific DLLs for different features and enables automatic version upgrades via future AMD Software: Adrenalin Edition releases.

Details:

  • Version Numbers: Includes AMD FSR Upscaling 4.0.3, AMD FSR Frame Generation 4.0.0, and AMD FSR Ray Regeneration 1.0.0.
  • Modular Architecture: The SDK now requires amd_fidelityfx_loader.dll as a base, with feature-specific libraries:
    • amd_fidelityfx_upscaler.dll: Provides FSR 4 (ML) and FSR 2/3.1.5 (Analytical).
    • amd_fidelityfx_framegeneration.dll: Provides FSR 4.0.0.
    • amd_fidelityfx_denoiser.dll: Provides Ray Regeneration 1.0.0.
    • amd_fidelityfx_radiancecache.dll: Provides Radiance Caching (Preview).
  • Driver-Level Upgrades: Games integrated with FSR 3.1+ or Frame Gen 3.1.4+ are eligible for automatic upgrades to FSR 4.x versions via the Adrenalin driver, decoupling the technology version from the game’s build.
  • Technical Requirements:
    • Requires Compute Shader 6.6 (CS_6_6) for certain features, necessitating DirectX 12 Agility SDK 1.4.9 or later.
    • ML-based features are optimized for RDNA 4 hardware (dedicated ML acceleration).
    • Backward compatibility for RDNA 3.5 and earlier is maintained via analytical paths (FSR 3 fallback).
  • Legacy Support: Previous FidelityFX technologies (pre-API based) remain in FidelityFX SDK 1.1.4. All new API-based features will live in SDK 2.0+.

[2025-12-10] AMD FSR™ Redstone for developers and the neural rendering future

Source: AMD GPUOpen

Key takeaway relevant to AMD: > AMD is rebranding the entire “FidelityFX” suite under the “AMD FSR” name to simplify developer integration and consumer recognition, while confirming that ML-based features are purpose-built for the upcoming Radeon RX 9000 series (RDNA 4).

Summary: > This report outlines the strategic shift toward neural rendering to overcome the scaling limitations of traditional rendering pipelines. It details the technical inner workings of ML-based Frame Generation and Radiance Caching, while announcing the deprecation of the “FidelityFX” brand in favor of a unified “AMD FSR” ecosystem.

Details:

  • Hardware Target: ML-powered features are specifically optimized for RDNA 4 architecture (Radeon RX 9000 series) using dedicated ML hardware.
  • ML Frame Generation Logic: Uses a neural approach taking previous/current frames, depth, and motion vectors. It combines optical flow estimation with an ML network that predicts per-pixel motion and appearance, significantly reducing ghosting and motion discontinuities compared to FSR 3.
  • Radiance Caching (NRC): Currently in “Technical Preview.” It uses an online ML model to train on-device in real-time to predict light propagation (Global Illumination). This replaces secondary ray bounces in path tracing with learned predictions.
  • Driver Requirement: The Radiance Caching preview requires AMD Software: Adrenalin Edition 25.12.1.
  • Unreal Engine Integration: Updated plugins are available for UE 5.2 through 5.7, including ML-based Frame Generation and Radeon Anti-Lag 2 support.
  • Branding Unified: “AMD FidelityFX” naming is being retired. All technologies (Upscaling, Frame Gen, Denoisers) will now be grouped under the AMD FSR brand.
  • Research Pipeline: AMD highlighted future research areas including Neural Supersampling and Denoising (NSSD), Generative AI for Global Illumination (GGI), and Dense Geometry Format (DGF).