Skip to navigation

Performance Results

View as MarkdownOpen in Claude

This page reports the measured inference speed of Relighting on supported GPUs.

Performance Data

NVIDIA AI for Media Relighting supports concurrent streams per GPU. The number of concurrent streams depends on GPU memory and video resolution.

The following table shows the performance data for Relighting for a single stream (concurrency 1) on various GPUs at four input video formats. The figures come from NVIDIA internal benchmark runs and are provided for information only. Actual results vary with hardware configuration, input characteristics, and workload. The GPU memory, host memory, and notebook figures elsewhere in this documentation were measured on an RTX 6000 Ada Generation GPU, which this table does not cover.

GPUInput VideoAverage FPS
RTX PRO 6000 Blackwell Server Edition720p, 30 fps68
1080p, 30 fps70
1080p, 60 fps73
4K, 30 fps55
L40S720p, 30 fps63
1080p, 30 fps63
1080p, 60 fps65
4K, 30 fps50
A10G720p, 30 fps32
1080p, 30 fps32
1080p, 60 fps32
4K, 30 fps27
L4720p, 30 fps25
1080p, 30 fps25
1080p, 60 fps25
4K, 30 fps22

The input videos are H.264 MP4 files: 720p at 30 fps (301 frames), 1080p at 30 fps (300 frames), 1080p at 60 fps (480 frames), and 4K at 30 fps (300 frames). Each request uses HDRI preset 0, the default background and gain settings, and lossy output at 10 Mbps. Each figure is the median of three benchmark runs, one request per run, after one warm-up request. The 720p and both 1080p inputs are streamable, so the NIM processed them in streaming mode; the 4K input is not, so it ran in transactional mode.

The inference FPS is calculated by dividing the total number of frames in the input video file by the total inference time in seconds (measured from the time the request is sent until the complete output file is received by the client).

For more information, refer to the NIM Clients GitHub repository: NVIDIA AI for Media NIM Clients.