NVIDIA VA-API driver, the VA-API hardware acceleration implemention using NVIDIA GPU, released new 0.0.17 version a few days ago.
The new version of this free open-source library fixed only two but important bugs for JPEG decoding and NVIDIA DGX Spark memory systems.
Firefox Web Content acceleration through NVIDIA GPU
As you may know, Linux has several ways (VA-API, VDPAU, AMF, NVDEC, or Vulkan Video) for decoding video content with GPU hardware acceleration. Firefox, the most popular and default browser in many Linux Distributions, however only supports VA-API so far in Linux.
Both Intel and AMD GPUs support VA-API, but NVIDIA does NOT. If you’re Firefox user but running Linux on computer with NVIDIA GPU, then the web content will be always decoded by your CPU by default.
To workaround the issue, the free open-source NVIDIA VA-API driver was born.
JPEG Decoding support
The driver previously supports for decoding AV1, H.264, HEVC, VP8, VP9, MPEG-2, VC-1 codecs. With the new 0.0.17, JPEG decoding is also supported.
The JPEG codec was previously unusable due to a structural mismatch. VI-API outputs decoded data as separate buffers (e.g., picture params, quantization tables, Huffman tables, slice params/data), but NVIDIA NVDEC expects a fully formed JPEG bitstream.
To fix this mismatch, NVIDIA VA-API driver since 0.0.17 now creates smallest valid JPEG file that conforms to the JFIF standard, directly from the VA-API buffers, before handing it off to NVDEC.
According to the feature request, the results are exactly identical at the byte level and hash-verified compare to the decoding outputs using the native NVDEC (CUDA). Though there are few limitations (see the previous link).
Fix for DGX Spark
NVIDIA DGX Spark, the desktop AI supercomputer, uses unified system memory shared between the CPU and GPU rather than discrete VRAM.
NVIDIA VA-API driver will now check twice during initialization to make sure it’s running on NVIDIA device with discrete VRAM or unified memory.
When unified memory is detected, it will remember the detection result, add required dependencies from open-gpu-kernel-modules, and use the correct new method to allocate memory. Thus, it will work properly on DGX Spark with Firefox.
Get NVIDIA VA-API Driver 0.0.17
The source tarball as well as the release note is available in Github page via the link below:
For Ubuntu users who want to try it out without building by yourself, I’ve made the new version into this unofficial PPA for Ubuntu 22.04, Ubuntu 24.04, Ubuntu 25.10 and Ubuntu 26.04.
If you don’t know how to get it work, see either the official README or follow this step by step guide.
