RT-Embed Performance | NVIDIA Blueprint for Video Search and Summarization

Overview

The Real-Time Embedding (RT-Embedding) microservice provides video and text embedding generation capabilities using the Cosmos-Embed1 model, enabling video search and similarity matching. Benchmarks cover two operating modes:

Streaming mode — the microservice reads live RTSP streams continuously. Every 10 seconds of video, it samples 8 frames and produces one video embedding vector. Latency is measured from when a 10-second chunk is ready until the embedding is returned.
Non-streaming mode — the microservice handles discrete embedding requests for video clips or text queries. Benchmarks measure throughput (requests per second) and latency across varying concurrency levels.

Each video embedding encodes 8 frames sampled from a 10-second chunk scaled to 448×448 resolution into a 768-dimensional vector. Text embeddings encode up to 128 tokens into the same 768-dimensional space, enabling cross-modal search.

RT-Embedding video embedding throughput vs concurrency for H100 and RTX Pro 6000 SE — Video embedding throughput (req/s) vs. concurrent requests. The H100 reaches up to 16 req/s while the RTX Pro 6000 SE peaks near 15 req/s; the RTX Pro 6000 SE reaches peak throughput at lower concurrency and saturates earlier, while the H100 sustains throughput through higher concurrency levels.

Test Configuration

Parameter	Value
VSS Release	3.2
Model	Cosmos-Embed1-448p
Model precision	FP16
Inference engine	TensorRT (TRT)
Frames per chunk (video)	8
Image resolution	448×448
Embedding vector length	768
Chunk duration	10 seconds
Text ISL (text embedding)	20 tokens
GPUs tested	H100, RTX Pro 6000 SE

Performance by GPU

H100

RTX Pro 6000 SE

Streaming Mode

Max Streams	Avg Latency (s)	p90 (s)	p95 (s)	Max (s)	GPU Core (%)	NVdec (%)
170	3.00	6.76	8.62	12.89	66.3	93.9

Concurrent Streams	Avg (s)	p90 (s)	p95 (s)	Min (s)	Max (s)	GPU Core (%)	NVdec (%)
1	0.13	0.13	0.19	0.08	0.32	0.0	0.6
16	0.20	0.20	0.25	0.08	0.29	4.0	6.4
32	0.24	0.24	0.31	0.08	0.34	8.0	11.7

Request Mode

Concurrent Requests	Avg Latency (s)	Throughput (req/s)
1	0.41	2.41
2	0.42	4.68
4	0.44	8.74
8	0.50	14.79
16	0.94	15.74
32	1.81	16.28

Concurrent Requests	Avg Latency (s)	Throughput (req/s)	p90 (s)	p95 (s)
1	0.007	132.1	0.007	0.008
2	0.008	216.7	0.009	0.011
4	0.010	368.0	0.010	0.012
8	0.013	563.6	0.015	0.016
16	0.018	800.5	0.023	0.023
32	0.027	1044.3	0.031	0.031

Video Duration	E2E Latency (s)	Avg Chunk Inference (s)	Decode Latency (s)	GPU Core (%)	NVdec (%)
10 s	0.412	0.039	0.369	0.0	0.0
10 min	5.203	0.039	0.615	48.9	72.6
60 min	28.8	0.039	0.601	54.3	87.6

All benchmarks use Cosmos-Embed1-448p, FP16, TensorRT, 8 frames per 10-second chunk of 1080p input resolution, scaled to 448×448 model resolution, and a 768-dimensional embedding vector. For streaming deployments, plan for 10–15% headroom below the maximum concurrent stream counts.