FFmpeg 9.0 “Lei”: Vulkan Expands, ONNX Hits GPU

FFmpeg 9.0 “Lei” shipped today — a major release of the open-source multimedia framework that quietly powers most of the video on the internet. The upgrade continues a clear trajectory: more Vulkan GPU acceleration, AI inference moving onto the GPU, and a couple of removals that will catch some pipelines off guard. Here is what changed and what you need to check before upgrading.

Vulkan Keeps Spreading Across the GPU Pipeline

The most consistent story in recent FFmpeg releases is Vulkan, and 9.0 is no different. The new v360_vulkan filter offloads 360-degree and panoramic video processing to any Vulkan-capable GPU — previously this was CPU-only and painfully slow for 8K spherical content. APV, Samsung’s open professional video codec backed by DaVinci Resolve and Android 16+, now has Vulkan hardware acceleration for both encoding and decoding. If you work in post-production on non-Apple hardware, this is a meaningful addition.

On the vendor-specific side: NVIDIA gets a transpose_cuda filter that keeps frame rotation on the GPU without copying frames back to system memory. AMD’s Advanced Media Framework picks up a hardware frame-rate converter (vf_frc_amf), a video quality enhancer (vf_vqe_amf), and broader HDR support in the color-conversion filter. Apple hardware gets VideoToolbox-accelerated ProRes RAW decoding. The picture forming here is that Vulkan is the cross-platform GPU path FFmpeg is betting on — vendor APIs fill in where Vulkan does not reach yet.

ONNX Runtime DNN Backend Now Runs on GPU

This one deserves more attention than it will likely get. FFmpeg’s DNN filter — which lets you run neural network models inside the filter graph — now supports ONNX Runtime GPU execution providers. That means AI upscaling, super-resolution, object detection, and background segmentation can run on GPU: NVIDIA CUDA, Windows DirectML across all GPU vendors, and AMD Ryzen AI NPUs via the VitisAI execution provider.

Put this alongside FFmpeg 8.0’s Whisper filter for automatic speech recognition, and a clearer picture emerges: FFmpeg is quietly becoming an AI-native media processing layer. You can now run a pipeline that decodes video on GPU, filters it through a Vulkan shader, runs an ONNX upscaling model on GPU, and outputs the result — without any frames touching the CPU. That is architecturally significant, even if the tooling to do it conveniently still requires some assembly.

Two Things Are Gone — Check Your Pipelines

Two removals in 9.0 warrant attention before you upgrade:

• CELT decoder removed. The standalone CELT decoder and Ogg/CELT parser are gone. Critically, this does not affect the Opus codec — Opus has its own internal CELT implementation and is completely unaffected. Standalone CELT predates Opus and was effectively obsolete since 2012. If you have no idea what CELT is, you are almost certainly fine.
• NVIDIA Video Codec SDK pre-11.1 support dropped. If your FFmpeg build links against an NVENC SDK older than version 11.1, GPU encoding will break after upgrading. Run ffmpeg -encoders | grep nvenc to check if NVENC is in your build and verify your SDK version before upgrading. Five years is a reasonable deprecation window, but it still bites people on long-running server builds.

Everything Else Worth Noting

Animated WebP finally gets native support: a dedicated decoder and demuxer land in 9.0, meaning FFmpeg can now read animated WebP files directly without reaching for libwebp or external tooling. LCEVC track muxing lands in the MP4 muxer — LCEVC is an enhancement layer that claims 40% bitrate savings on top of any base codec, making it relevant for live streaming and eSports workflows. HE-AAC 960 decoding adds DAB+ digital radio compatibility for European broadcast pipelines. A bitstream filter for splitting Dolby Vision multi-layer HEVC streams rounds out the professional broadcast additions. And yes, there is a Playdate video encoder and muxer, because open-source projects contain multitudes.

Should You Upgrade?

If you are not using NVENC with an old SDK build and have no legacy CELT streams, the answer is yes. The GPU acceleration improvements are real, the ONNX DNN on GPU is a forward-looking feature worth experimenting with, and animated WebP support will simplify a common pipeline annoyance. FFmpeg 9.0 is available now at ffmpeg.org. Check the Phoronix breakdown and Linuxiac’s feature list for the full changelog detail before you push to production.