Fastvideo SDK for GPU Image Processing

Updated: September 17, 2026

Fastvideo SDK is a C++/CUDA library that runs the complete camera image processing pipeline on NVIDIA GPUs — from raw image acquisition to JPEG or JPEG2000 compression and storage — without round-trips to the CPU. On an NVIDIA GeForce RTX 4090 the full RAW-to-RGB pipeline is so fast that end-to-end throughput is limited by the SSD, not the GPU: running entirely in GPU memory it processes a 65 MPix frame at about 60 fps, while a complete workflow that reads and writes to disk reaches up to 60 fps for 12-bit 4K images end-to-end. Individual modules run far higher — JPEG encoding at about 5500 fps (Full HD) or 3800 fps (4K), visually lossless JPEG2000 encoding at up to about 2100 fps (2K, batch mode), and demosaicing at up to 69 GPix/s. Frames are uploaded once, processed end-to-end in GPU memory, and the results are ready for display, streaming, or writing to disk in real time.

How these numbers were measured. Fastvideo benchmarks are measured on NVIDIA GPUs, timing pure GPU processing at the resolution, bit depth, format and quality stated on this page. See the full benchmark methodology and test environment and the Fastvideo SDK benchmark report (PDF). Our tools are open source on GitHub and built on NVIDIA CUDA.

Check a single module. Every module in the Fastvideo SDK comes with a command-line sample. Run it on your own images on your own GPU and compare the result with the figures published here.

Check the whole pipeline. A module in the middle of a pipeline cannot be judged on its own: its input is hard to prepare, and its output is an intermediate product you cannot look at. What matters is how it changes the final image and the total time of the schema. Two of our applications run the complete pipeline and report the time of every stage — FastVCR (live cameras or RAW files from disk, so no camera is required) and Fast CinemaDNG (DNG files, offline), both built on our GPU RAW Processor pipeline. The benchmark module is included in the free builds of both applications.

Detailed benchmark data. Measured results are published on our benchmark pages: all benchmarks.

Want the internals? See how the modules chain in GPU memory on the CUDA image processing library architecture page. This page is the Fastvideo SDK product — features, benchmarks, and licensing.

The SDK is built for applications where both throughput and image quality matter: machine vision and industrial cameras, digital cinema, broadcasting, medical imaging, aerial, aerospace and street-view imaging, and embedded systems on NVIDIA Jetson. Beyond camera pipelines, the same components power offline RAW processing, FFmpeg codecs and filters, and preprocessing for AI applications.

Fastvideo SDK image processing pipeline on NVIDIA CUDA: camera, framegrabber, SDK, output — The Fastvideo SDK image processing pipeline on NVIDIA CUDA: frames flow from the camera through a framegrabber into the SDK — preprocessing, demosaicing, color grading, resize, and JPEG/JPEG2000/H.264 encoding on the GPU — and out to display, storage, or network.

Fastvideo SDK at a glance

What it is: C++ library with CUDA backend, ships with sample applications and source code, API, and documentation
Pipeline: full ISP from RAW/Bayer to compressed output, entirely on GPU
Input sources: camera, frame grabber, HDD/SSD/RAM or GPU memory
Input formats: PGM, BMP, PPM, JPG, YCbCr, RAW, DNG, byte array, OpenGL texture/PBO
Output: display via OpenGL, JPEG and JPEG2000 compressed streams, AVI/MXF containers
Platforms: Windows 10/11, Linux Ubuntu, Linux4Tegra; NVIDIA GPUs from mobile to server, including Jetson
Performance: real-time processing of multi-camera pipelines at resolutions from 4K up to 12K and beyond

Fastvideo SDK performance benchmarks

How fast are the Fastvideo SDK modules on a GPU?

Representative peak throughput of the main SDK modules on an NVIDIA GeForce RTX 4090 (pure GPU processing time, 24-bit color; the mode is noted per row). Exact numbers depend on your GPU, resolution and quality — full per-module results are on the linked benchmark pages.

Representative per-module performance of Fastvideo SDK on the NVIDIA GeForce RTX 4090, by operation and image resolution.
Module	Operation	Resolution	Speed on RTX 4090
RAW→RGB pipeline	Min pipeline, in GPU memory	65 MPix	60 fps
Full RAW→RGB pipeline	End-to-end, incl. SSD read/write	4K, 12-bit	60 fps
JPEG codec	Encoding	Full HD / 4K / 6K	5500 / 3800 / 1500 fps
JPEG codec	Decoding	4K / 8K	1350 / 650 fps
JPEG2000 codec	Encoding (batch, visually lossless)	2K / 4K	2108 / 732 fps
JPEG2000 codec	Decoding (batch, visually lossless)	2K / 4K	1317 / 362 fps
Debayer / demosaic	HQLI … MG algorithms	peak throughput	20 - 69 GPix/s

The RAW→RGB rows are the full ISP pipeline: the in-memory case is GPU-bound, while the end-to-end case is limited by the SSD, not the GPU. JPEG figures are single-thread single-image; JPEG2000 figures are for multithreaded batch mode — single-image results and full tables are on the JPEG2000 encoder and decoder benchmark pages. The JPEG codec reaches about 117 GB/s on a 24 MPix (6K) frame. See detailed benchmarks for every module.

Why Fastvideo SDK

Why run the full image processing pipeline on the GPU?

The whole pipeline on the GPU, not just parts of it. Individual GPU-accelerated functions lose their advantage if the data has to travel back to the CPU between stages. Fastvideo SDK keeps the entire pipeline in GPU memory, so the PCIe bus is crossed only twice — raw data in, final result out. This is what makes real-time 4K and multi-camera processing achievable in practice.

Performance without sacrificing image quality. Every module is designed to hold its own against the best CPU implementations in quality — including Bilateral and NLM denoising, DFPD and MG demosaicing, and Lanczos resizing — while running significantly faster. Published benchmarks for every major module back this up.

One SDK across the full NVIDIA range. The same API runs on everything from Jetson modules to laptop, desktop, and server GPUs. An application developed on a workstation deploys to an embedded Jetson system without code changes — only the performance envelope changes.

Lower total cost of the final application. GPU processing offloads the CPU, so a single machine handles workloads that would otherwise require several — fewer computers per camera system, less rack space, simpler maintenance. The SDK is compatible with third-party CPU and GPU libraries, so it integrates into existing applications instead of replacing them.

We wrote these algorithms ourselves

Image processing has always been a trade-off between quality and speed. The methods that give a good result take minutes per frame; the fast ones leave halos and artifacts. The reason is the compute budget: on a CPU there was never enough of it, so algorithms were simplified — and much of the industry still runs on simplifications accepted decades ago.

Fastvideo’s codecs and ISP modules are neither ports of that legacy nor wrappers around third-party code. We designed — and in many cases invented — the algorithms behind our demosaicing, denoising, tone mapping and JPEG/JPEG2000 codecs, and wrote them for CUDA from the start, on top of image-processing expertise built since 2009. An algorithm moved to the GPU still carries the compromises it was simplified for; ours were never made under that constraint. On a modern GPU the budget is large enough that, for practical workloads, the choice between quality and speed no longer has to be made.

That is why you get a complete, tuned imaging solution rather than a bag of GPU primitives — and why, when your pipeline needs something specific, we can change the algorithm itself, not just call it.

Trusted by imaging teams

“Fastvideo SDK is lightning fast… we can now process a 28,800 image sequence set in under 10 minutes, instead of 10 hours.”
— Infinite Realities

Image & Video Processing SDK Features

Image acquisition

Fastvideo works as an image acquisition SDK: frames from any source go straight into GPU memory for processing.

Image ingestion from a camera, frame grabber, HDD/SSD/RAM or GPU memory
Input formats: PGM, BMP, PPM, JPG, YCbCr, RAW, DNG, byte array, OpenGL texture or PBO

RAW domain processing

RAW data linearization
Dark frame subtraction (FPN)
Flat-field correction (shading correction)
Dynamic bad pixel correction
Binning for grayscale images
White balance
1D LUT for each bayer channel
RAW histograms and RAW Bayer curves (master and RGB)
RAW denoising — in development
Chromatic aberration removal in the RAW domain — in development

Demosaicing

Algorithms: HQLI, L7, DFPD, MG
Demosaicing with binning (B2, B4, B8)
After-demosaicing filter — in development

Color and tone

Exposure correction and highlight recovery
Color correction with matrix profile
Color transforms: RGB↔YCbCr, RGB↔HSV, and more
Composite or per-channel 8/10/12/14/16-bit 1D LUTs for tone mapping in RGB or HSV (Hue→Hue, Hue→Saturation, etc.)
3D LUT for RGB and HSV
Gamma transform (composite or per-channel)
Transforms to Rec.601 (SD), Rec.709 (HD), Rec.2020 (4K) for broadcasting

Denoising and enhancement

Denoising: bilateral, NLM
Sharpening with Unsharp Mask filter (USM)
Defringe
Tone mapping with ALTMapper algorithm — in development
Fog removal and dehaze — in development

Geometric transforms

Crop, resize (downscale or upscale)
Rotate to 90/180/270 degrees, flip/flop
Standard remap: rotation to an arbitrary angle, affine, perspective, projection, arbitrary warping

Analysis and monitoring

RGB histograms
RGB parade
Waveform monitor

Output and codecs

Real-time display output via OpenGL
JPEG codec (8/12 bits per channel)
JPEG2000 encoder and decoder (8–16 bits per channel, tiles, batch, multithreading, multithreaded batch)

Pipeline infrastructure

Stream-per-thread support for better performance
CUDA MPS support for Linux
FFmpeg filters and codecs on GPU for high performance video processing

SDK Options

The following modules are also part of the SDK. They target less common tasks and are licensed on demand.

Extended format support

Unpacking module for various RAW formats
RAW and DNG (CinemaDNG) image and video processing
MLV raw video processing
OME-TIFF, BigTIFF, GeoTIFF support (tiles, pyramids, encoding)
HDR for high bit depth image sensors

Video codecs

H.264 codec (NVENC and NVDEC)
H.265 codec (NVENC and NVDEC)
J2K codec on GPU for FFmpeg

I/O and integration

Input from SD/HD-SDI and 3G/6G/12G-SDI grabbers: Blackmagic, Imperx
Multiple camera solutions for real-time image processing

Compatibility

OS: Windows 10/11, Linux Ubuntu, Linux4Tegra (L4T)
NVIDIA GPUs: Pascal, Volta, Turing, Ampere, and Ada generations — mobile, laptop, desktop, and server models; support for the latest GPU generations is in progress
NVIDIA Jetson: Nano, Xavier NX/AGX, Orin NX/AGX
CUDA 12.6; MSVS 2022 on Windows
Ships with sample C++ applications with source code, video processing API, and manual
Motion JPEG integration with FFmpeg (reader/writer); FFmpeg is under LGPL v2.1

SDK Sample Applications with Source Code

CUDA JPEG codec
Debayer, and Debayer + JPEG
Denoiser
Resizer
J2K codec
SDI converter
JPEG Resize (Jpeg2Jpeg): JPEG decoding, crop, resize, sharp, JPEG encoding
Camera application in C++ and CUDA: raw data acquisition and unpacking, dark frame subtraction, shading correction, white balance, exposure correction, debayer, luma and chroma denoising, color correction, tone mapping with LUT (curves and levels), rotate, crop, resize, remap, undistortion, sharp, OpenGL output, JPEG or JPEG2000 compression, writer to AVI (MJPEG) or MXF (JPEG2000)
FFmpeg Remap filter
FFmpeg J2K decoder

Open Source Projects from Fastvideo on GitHub

PGM2DNG converter — converts PGM images to DNG
GPU Camera Sample — software for machine vision and industrial cameras with GPU image processing; Windows binaries are available from the same repository

GPU-Based Demo Applications for Download

Licensing and Custom Development

We license the GPU Image & Video Processing SDK and its components for Windows, Linux, and L4T to software developers, camera manufacturers, system integrators, and resellers. The SDK is used in a wide range of real-time imaging applications, from machine vision to digital cinema and broadcasting.

Fastvideo also develops custom image and video processing solutions according to an agreed specification, and offers integration services to connect the SDK with your existing software. If your image processing pipeline includes a stage that is not listed on this page, contact us to discuss it — you can also check our projects in development.

Frequently asked questions

What is Fastvideo SDK?

Fastvideo SDK is a C++/CUDA library that runs the complete camera image processing pipeline on NVIDIA GPUs — from raw acquisition, demosaicing, denoising, color correction and tone mapping, and resize to JPEG and JPEG2000 compression — entirely in GPU memory.

How fast is it?

On an NVIDIA GeForce RTX 4090 the full RAW-to-RGB pipeline processes a 65 MPix frame at about 60 fps in GPU memory (up to 60 fps end-to-end for 12-bit 4K, where throughput is limited by the SSD, not the GPU). Individual modules run higher still: JPEG encoding at about 5500 fps (Full HD) and 3800 fps (4K), visually lossless JPEG2000 encoding at up to about 2100 fps (2K, batch mode), and demosaicing at up to 69 GPix/s — see the benchmark methodology.

Which GPUs and platforms are supported?

Any CUDA-capable NVIDIA GPU, from Jetson embedded modules (Xavier, Orin) to laptop, desktop, and server cards, on Windows 10/11, Linux Ubuntu, and Linux4Tegra. The same C++/CUDA API runs across the whole range without code changes.

Which pipeline stages does the SDK cover?

RAW-domain preprocessing, demosaicing, denoising, color correction and tone mapping, geometric transforms, analysis and monitoring, and output codecs — JPEG, JPEG2000, and H.264/H.265 via NVENC/NVDEC. It also integrates FFmpeg filters and codecs on the GPU.

Can I test the performance myself?

Yes. Download the free demo apps for individual modules and measure the speed on your own GPU — see the benchmark methodology and the Fastvideo SDK benchmark report (PDF).

Is the Fastvideo SDK an image acquisition SDK?

Yes. The SDK covers the full image-acquisition path — ingesting frames from a camera, frame grabber, SSD/RAM or GPU memory in formats such as PGM, RAW, DNG and OpenGL textures — and then runs the complete ISP (demosaicing, color, denoising, tone mapping) and JPEG/JPEG2000 encoding on the GPU, so image acquisition and processing stay in one GPU pipeline.

How is the Fastvideo SDK different from NVIDIA NPP?

NPP gives you low-level GPU primitives that you assemble into a pipeline yourself. The Fastvideo SDK is a ready end-to-end camera ISP — RAW-to-RGB (demosaicing, color, denoising, tone) plus JPEG/JPEG2000 output — that runs the whole pipeline in GPU memory in up to 16-bit precision, so you link one library instead of building the pipeline from parts. See the detailed comparison in Fastvideo SDK vs NVIDIA NPP.

Roadmap for CUDA Image Processing SDK

Fastvideo SDK is a mature product, developed and refined since 2009; its latest releases (2026) add support for the newest NVIDIA GPUs and CUDA versions. The items below are the most recent additions and the work currently in progress.

Released:

Latency improvement for GPU-based ISP
High quality color calibration software
CUDA 12.6 support; CUDA 12.2 support for Jetson Orin
FastVCR software for XIMEA cameras (real-time raw processing on Jetson with integrated camera control)

In progress:

ALTMapper (Adaptive Local Tone Mapping) engine for single shot HDR
Chromatic aberration removal in the RAW domain with self-calibration
RawZipper — Raw Bayer codec on GPU for lossy compression
CUDA 13.3 support
Jetson Thor support
Accelerated Remap
GPU memory manager

About the author

Fyodor Serzhenko, PhD, is the founder and CEO of Fastvideo. He earned his PhD at the Moscow Institute of Physics and Technology (MIPT) in 1993. Since 2009 he has led the development of Fastvideo’s GPU-accelerated image codecs and ISP modules, including the Fastvideo SDK. Connect on LinkedIn.

Why you can trust these results

Fastvideo has built GPU-accelerated image processing software since 2009. Our codecs comply with the standards they implement — JPEG (ITU-T T.81 / ISO IEC 10918) and JPEG2000 (ITU-T T.800 / ISO IEC 15444) — and our tools are open source on GitHub. Every performance figure we publish is reproducible: see our benchmark methodology, download the Fastvideo SDK benchmark report (PDF), and measure it on your own GPU.