AIDP Video Forge

GPU-Accelerated Video Processing on Decentralized Compute Networks

Matthew Karsten · Purple Squirrel Networks · February 2026

gpu-acceleration nvenc cuda video-processing

Abstract

We present AIDP Video Forge, a GPU-accelerated video processing system leveraging decentralized compute networks. Our approach utilizes NVIDIA hardware encoding (NVENC) and CUDA-accelerated filters across distributed GPU nodes to provide 10-20x faster video encoding compared to CPU-based methods. Through intelligent job orchestration and distributed batch processing, we achieve 40-60% cost reduction versus centralized cloud GPU services while maintaining professional-grade video quality.

Key Results

16x

Faster than CPU

58%

Cost Reduction

95.8

VMAF Score

37x

Distributed (5 GPU)

Metric	AIDP Video Forge	AWS MediaConvert	Improvement
Encoding Speed (4K)	2.8 min	3.2 min	16x faster than CPU
Cost per Hour	$0.25	$0.60	58% cheaper
Quality (VMAF)	95.8	96.0	Near-identical
Distributed (5 GPUs)	1.2 min	N/A	37x faster than CPU

Architecture

+----------------------------------------------------------+
|                     Video Forge                          |
+----------------------------------------------------------+
|  Client (Web UI / CLI)                                   |
|  +-- Upload video -> Select processing -> Download       |
+----------------------------------------------------------+
|  Job Orchestrator                                        |
|  +-- Queue jobs -> Assign to AIDP nodes -> Aggregate     |
+----------------------------------------------------------+
|  AIDP GPU Workers                                        |
|  +-- FFmpeg + NVENC + CUDA filters                       |
+----------------------------------------------------------+

GPU Acceleration: NVENC vs CPU

Operation	CPU Method	GPU Method	Speedup
H.264 Encoding	libx264	h264_nvenc	15-20x
HEVC Encoding	libx265	hevc_nvenc	20-30x
Scaling	scale	scale_cuda	5-8x
Deinterlacing	yadif	yadif_cuda	8-10x
HDR Tone Map	zscale+tonemap	tonemap_cuda	15x
LUT Application	lut3d	CUDA texture	10x

Processing Speed Benchmark

Method	Time (10-min 4K)	Real-time Speed	Speedup
CPU (libx264)	45 minutes	0.22x	1x baseline
AWS MediaConvert (T4)	3.2 minutes	3.1x	14x faster
AIDP Video Forge (RTX 3090)	2.8 minutes	3.6x	16x faster
Distributed (5 GPUs)	1.2 minutes	8.3x	37x faster

Technical Contributions

Hardware Acceleration: Full NVENC/CUDA pipeline eliminating CPU bottlenecks
Distributed Processing: Intelligent job splitting across multiple GPU nodes
Cost Efficiency: 40-60% reduction vs centralized cloud GPU services
Quality Preservation: VMAF 95.8 — near-identical to reference encoding

Citation

@techreport{karsten2026videoforge, title={AIDP Video Forge: GPU-Accelerated Video Processing on Decentralized Compute Networks}, author={Karsten, Matthew}, institution={Purple Squirrel Networks}, year={2026}, month={February}, url={https://huggingface.co/purplesquirrelnetworks/aidp-video-forge-paper} }