Nvidia NVENC

The first generation of NVENC, which is shared by all Kepler-based GPUs, supports H.264 high-profile (YUV420, I/P/B frames, CAVLC/CABAC), H.264 SVC Temporal Encode VCE, and Display Encode Mode (DEM).

NVidia's documentation states a peak encoder throughput of 8× realtime at a resolution of 1920×1080 (where the baseline "1×" equals 30 Hz). Actual throughput varies on the selected preset, user-controlled parameters and settings, and the GPU/memory clock frequencies. The published 8× rating is achievable with the NVENC high-performance preset, which sacrifices compression efficiency and quality for encoder throughput. The high-quality preset is considerably slower but produces fewer compression artifacts.

Introduced with the first-generation Maxwell architecture, second generation NVENC adds support for the high-performance HP444 profile (YUV4:4:4, predictive lossless encoding), and increases encoder throughput up to 16× realtime, which corresponds to about 1080p @ 480 Hz with the high-performance preset.)

Maxwell GM108 does not have NVENC hardware encoder support.

Introduced with the second-generation Maxwell architecture, third generation NVENC implements the video compression algorithm High Efficiency Video Coding (a.k.a. HEVC, H.265) and also increases the H.264 encoder's throughput to cover 4K-resolution at 60 Hz (2160p60). However, it does not support B-frames for HEVC encoding (just I and P frames). The maximum NVENC HEVC coding tree unit (CU) size is 32 (the HEVC standard allows a maximum of 64), and its minimum CU size is 8.

HEVC encoding also lacks Sample Adaptive Offset (SAO). Adaptive quantization, look-ahead rate control, adaptive B-frames (H.264 only) and adaptive GOP features were added with the release of Nvidia Video Codec SDK 7. [11] These features rely on CUDA cores for hardware acceleration.

SDK 7 supports two forms of adaptive quantization; Spatial AQ (H.264 and HEVC) and Temporal AQ (H.264 only).

Nvidia's consumer-grade (GeForce) cards and some of its lower-end professional Quadro cards are restricted to three simultaneous encoding jobs. Its higher-end Quadro cards do not have this restriction.

Fourth generation NVENC implements HEVC Main10 10-bit hardware encoding. It also doubles the encoding performance of 4K H.264 & HEVC when compared to previous generation NVENC. It supports HEVC 8K, 4:4:4 chroma subsampling, lossless encoding, and sample adaptive offset (SAO).

Nvidia Video Codec SDK 8 added Pascal exclusive Weighted Prediction feature (CUDA based). Weighted prediction is not supported if the encode session is configured with B frames (H.264).

There is no B-Frame support for HEVC encoding, and the maximum CU size is 32×32.

The NVIDIA GT 1030 and the Mobile Quadro P500 are GP108 chips that don't support the NVENC encoder.

In laptop graphics, NVIDIA MX Graphics do not include NVENC as they are based on a Maxwell-generation GM108 or a Pascal-generation GP108 chip. [12] The GeForce MX350 is a GP107 chip whose NVENC encoder is disabled during manufacture.

Volta NVENC has similar performance as Pascal's NVENC.

It does not offer support for HEVC B-Frames.

Sixth generation NVENC implements HEVC 8K encoding at 30FPS, HEVC B-Frames support and provides up to 25% bitrate savings for HEVC and up to 15% bitrate savings for H.264. The Nvidia GeForce GTX 1650 is exempt from this generation however, as it uses Volta NVENC instead of Turing. The GTX 1650 Super however uses the Turing NVENC engine as it is based on the TU116 rather than the TU117 used in the regular GTX 1650. [13]

Ampere has essentially the same NVENC generation engine as Turing.[14]