Performance Reference

This table lists latency and throughput data for features of the NVIDIA Video Effects SDK on a few supported GPU architectures. Lower latency and higher throughput are better.

	Latency (in milliseconds)				Throughput (max streams to achieve 30 FPS)
Feature	T4	A40	L40	B40	T4	A40	L40	B40
Video Super Resolution, Streaming Medium (mode 21)	N/A	1.52	0.806	0.583	N/A	21	41	57
Video Super Resolution, Streaming Ultra (mode 23)	N/A	4.74	3.14	1.60	N/A	7	10	20
Video Frame Generation, Medium model, 2× frame multiplier	N/A	N/A	1.29	1.15	N/A	N/A	25	28
TrueHDR, debanding enabled	1.16	0.451	0.179	0.167	28	73	186	199
Webcam Denoising	2.51	1.03	0.491	0.367	13	32	67	90
Upscale	0.53	0.132	0.0788	0.0471	62	252	422	707
AI Green Screen	4.11	1.55	0.762	0.849	8	21	43	39
Video Relighting	50.5	15.6	7.49	5.01	0	2	4	6

Note

• Unless otherwise specified, effects were tested at 1280 × 720 input and output resolution.

• Video Super Resolution used 1280 × 720 input and 2560 × 1440 output (2× scaling in each dimension).

• Video Frame Generation used 1280 × 720 input and output and generated the intermediate frame at t=0.5.

• TrueHDR used 1280 × 720 input and output with DebandingOff=0, the default, which enables debanding.

• Upscale used 1280 × 720 input and 1920 × 1080 output (1.5× scaling in each dimension).

• Throughput values are calculated from per-call latency as the maximum integer number of streams that can achieve 30 FPS. The calculation uses the source latency before display rounding and is equivalent to floor(1000 / (latency_ms * 30)). These values are estimates and are not measured concurrent-stream results.

• N/A indicates an unsupported configuration or a GPU for which no benchmark result is available. The Video Super Resolution streaming modes and Video Frame Generation are unsupported on T4, and Video Frame Generation is also unsupported on A40.