Evaluation of Hardware Encoders on GPUs Demonstrates Performance for 4K/8K UHD Live-Streaming

The increasing popularity of live video content, from gaming streams to virtual events, places significant demands on efficient video encoding technology. Kasidis Arunruangsirilert and Jiro Katto, from Waseda University, along with their colleagues, investigate the performance of dedicated video encoders now integrated into modern graphics processing units (GPUs). Their work evaluates how these hardware-based encoders, found in GPUs from several manufacturers including Intel and Qualcomm, compare to traditional software encoding methods, particularly when handling ultra-high definition (UHD) video. The team assesses encoding speed, video quality, and power consumption, revealing that current GPU encoders achieve comparable results to software alternatives in real-time applications, though improvements between generations offer limited gains in efficiency. This research establishes a benchmark for evaluating hardware encoders and quantifies the bitrate needed to meet the quality standards of platforms like YouTube, providing valuable insight for developers and content creators.

GPU Encoders Beat Software for Efficiency

This research comprehensively evaluates hardware encoders in modern GPUs, Intel Quick Sync, NVIDIA NVENC, and Qualcomm Snapdragon, assessing their rate-distortion performance, encoding throughput, and power consumption. The study offers practical guidance for content creators, particularly streamers and VTubers, on optimizing encoding settings for high-quality video delivery. Key findings demonstrate that hardware encoders outperform software encoders, achieving comparable encoding speeds while consuming significantly less power. Intel Quick Sync generally delivers the best rate-distortion performance, requiring less bitrate to achieve the same quality compared to NVIDIA NVENC and Qualcomm Snapdragon.

NVIDIA NVENC performs slightly behind Intel Quick Sync, while Qualcomm Snapdragon’s hardware encoders significantly underperform, requiring 50% more bitrate to match the quality of Intel and NVIDIA. Improvements in rate-distortion performance are primarily driven by the adoption of newer codecs, such as AV1, rather than hardware improvements over the past six years. The study recommends using the newest codec supported by your hardware for live streaming to maximize quality, as YouTube’s H. 264/AVC encoding yields lower quality than VP9 and AV1. Researchers evaluated encoders using objective metrics, PSNR, SSIM, and VMAF, employing ultra-high definition test sequences and real-world content from Twitch.

They tested modern GPUs from Intel, NVIDIA, and Qualcomm, utilizing codecs including H. 264/AVC, H. 265/HEVC, and AV1, with various presets and resolutions to simulate real-world scenarios. The authors suggest future research should focus on evaluating newer encoding features, such as NVIDIA’s Split-Frame Encoding, and investigating encoding performance at higher resolutions, 8K, and frame rates. Incorporating subjective evaluations, user perception of quality, to complement objective metrics would also be beneficial. Overall, this paper provides a valuable resource for content creators seeking to optimize their video encoding settings for the best possible quality and efficiency.

Video Encoder Performance, Metrics and Comparisons

This study meticulously evaluated the performance of modern video encoders, both hardware-based in GPUs and software running on CPUs, to address the increasing demand for efficient encoding of high-resolution content, such as live streams and 4K/8K video. Researchers compared encoders across key metrics, rate-distortion performance, encoding speed, and power consumption, using a comprehensive methodology. To establish a baseline, the team encoded a diverse set of videos using the latest codecs, including H. 266/VVC, and assessed their quality using objective measures, Peak Signal-to-Noise Ratio, Structural Similarity Index Measure, and a machine-learning based metric called Video Multi-method Assessment Fusion.

The study incorporated a rigorous comparison against lossless downscaled versions of videos uploaded to YouTube, eliminating potential quality degradation from upscaling. The team analyzed performance across three distinct bitrate ranges, low, medium, and high, using multiple data points within each range to calculate average performance. Recognizing that benchmark datasets can be overly complex, researchers expanded their evaluation to include ten different types of commonly found video content, ranging from 1080p game cutscenes and Japanese animations to UHD documentaries and live sports broadcasts. These diverse sources were encoded using the AV1 codec on an Intel Arc A770 GPU, and the resulting rate-distortion curves were compared to those generated from the benchmark datasets. To determine realistic encoding settings for software encoders, the team tested various presets on a laptop with a high-end CPU, identifying the fastest preset that could maintain real-time encoding at 60 frames per second.

Real-Time Video Encoding Performance Benchmarks

This work presents a comprehensive evaluation of modern video encoders, both hardware-accelerated and software-based, focusing on real-time encoding performance for online streaming applications. Researchers meticulously assessed encoding efficiency, speed, and power consumption across several codecs, H. 264/AVC, H. 265/HEVC, VP9, and AV1, using a variety of content types at resolutions of 1080p and 4K. The study demonstrates that modern GPU hardware encoders now achieve rate-distortion performance comparable to software encoders when used in real-time encoding scenarios, effectively matching medium-quality presets.

Experiments revealed that while newer hardware generations generally exhibit increased encoding speed, improvements in rate-distortion performance between generations are negligible. To quantify performance, the team measured encoding quality using metrics including PSNR, SSIM, and a machine-learning based VMAF, alongside direct comparisons to YouTube’s transcoding quality. Results show that the Intel Arc A770 GPU consistently delivers the best rate-distortion performance among the tested GPU encoders. A key achievement of this research is the determination of the bitrate required for each hardware encoder to match YouTube’s transcoding quality. This data provides practical guidance for gamers, streamers, and VTubers seeking to optimize their live encoding settings. Furthermore, the study investigated the performance of each encoder across diverse content genres, including VTuber promotional videos, Japanese animation, film, and documentaries, both in Standard Dynamic Range and High Dynamic Range.

Hardware Encoders Outperform Software at Ultra HD

This research presents a comprehensive evaluation of hardware and software video encoders commonly found in modern GPUs, integrated graphics, and mobile SoCs. The team demonstrated that while software encoders struggle to deliver sufficient encoding throughput at ultra-high resolutions, modern hardware encoders achieve comparable video quality in real-time encoding scenarios. Analysis across multiple generations of hardware from NVIDIA, Intel, and Qualcomm reveals that improvements in encoding speed have been made, but gains in raw encoding performance have been limited. The study extends beyond simple performance metrics by assessing power consumption and calculating the bitrate required to match the quality of YouTube transcoding.