Performance

This section provides performance benchmarks for VSS across several video lengths and chunk sizes on an 8xH100 PCIE system with 4 GPUs used for the LLM and 4 being shared between the Embedding, ReRanker and VLM.

The summarization latency metrics for summarization are measured with Graph RAG enabled and disabled. If Graph RAG is enabled, it allows for interactive Q&A after the summary has been generated but adds latency.

Video Length (Minutes)	Chunk Size (Seconds)	Chunks	Summarization Latency - Graph RAG Off (Seconds)	Summarization Latency - Graph RAG on (Seconds)
1 min	10s	6	12.94	28.56
1 min	20s	3	10.60	24.19
1 min	30s	2	11.15	25.26
10 min	10s	60	55.25	89.88
10 min	20s	30	30.73	59.99
10 min	30s	20	28.17	51.52
30 min	10s	180	75.00	197.89
30 min	20s	90	48.67	116.96
30 min	30s	60	43.81	85.42
60 min	10s	360	138.86	382.25
60 min	20s	180	76.38	194.90
60 min	30s	120	67.59	141.41
120 min	10s	720	204.12	684.58
120 min	20s	360	126.31	364.79
120 min	30s	240	87.11	320.88

From the data, the summarization speed up can be plotted. This is summary generation time compared to the length of the video to show how much faster VSS is to generate summaries compared to manually watching the full video.

The following graph shows the relationship between latency, chunk size and video size. Longer videos and shorter chunk sizes require more processing and have higher latency.