> This page is for version Dev.
> For other versions, use one of these documentation indexes:
> - 3.3.0 (default): https://docs.nvidia.com/vss/3.3.0/llms.txt
> - Dev: https://docs.nvidia.com/vss/dev/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/vss/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/vss/_mcp/server.

# Video Summarization Performance

## Overview

The Video Summarization microservice processes uploaded video files through
[RTVI-VLM](/system-components/vision-microservices/real-time-vlm) for per-chunk vision inference and a summarization LLM (CA-RAG) for
event aggregation. The tables below report mean end-to-end (E2E) latency for file summarization
and maximum concurrent request capacity (burst) across video lengths. Use these numbers to size
GPU count and topology for batch summarization workloads on H100, RTX Pro 6000 SE, and L40S
platforms.

## Test Configuration

| Parameter           | Value                                                     |
| ------------------- | --------------------------------------------------------- |
| VSS release         | 3.2                                                       |
| VLM                 | Cosmos3 Nano Reasoner                                     |
| Model precision     | FP8                                                       |
| Summarization LLM   | Nemotron 3 Nano (default Video Summarization stack)       |
| Chunk duration      | 10 seconds                                                |
| Frames per chunk    | 20                                                        |
| Vision input tokens | 9k                                                        |
| Output max tokens   | 100                                                       |
| Test scenario       | Warehouse safety monitoring                               |
| Test videos         | 1, 10, 30, 60, 120, and 720 minute MP4                    |
| GPU topologies      | `1` (1 GPU), `2x2` (2 VLM + 2 LLM), `4x4` (4 VLM + 4 LLM) |
| Platforms tested    | H100, RTX Pro 6000 SE, L40S                               |

## E2E Latency

#### H100

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 5.16  | 20.1   | 52.9   | 64.7   | 119     |

720-minute E2E Avg: **615 s**

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 5.70  | 23.3   | 59.7   | 128    | 205     |

720-minute E2E Avg: **1151 s**

**Topology `1`** (1 GPU — VLM and LLM colocated)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 12.1  | 50.7   | 117    | 265    | 438     |

#### RTX Pro 6000 SE

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 6.50  | 25.9   | 51.6   | 95.2   | 187     |

720-minute E2E Avg: **935 s**

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 7.05  | 33.0   | 91.6   | 184    | 334     |

720-minute E2E Avg: **1805 s**

**Topology `1`** (1 GPU — VLM and LLM colocated)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 17.6  | 71.7   | 170    | 318    | 623     |

#### L40S

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 8.99  | 39.1   | 91.4   | 164    | 330     |

720-minute E2E Avg: **1890 s**

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length | 1 min | 10 min | 30 min | 60 min | 120 min |
| ------------ | ----- | ------ | ------ | ------ | ------- |
| E2E Avg (s)  | 12.2  | 65.8   | 168    | 340    | 609     |

720-minute E2E Avg: **3957 s**

## Maximum Concurrency

Maximum concurrent summarize requests at each target mean E2E latency. Test videos are 1, 5, and
10 minute MP4.

#### H100

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 73    | 114   | 125    |

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 42    | 65    | 72     |

#### RTX Pro 6000 SE

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 41    | 71    | 79     |

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 22    | 38    | 41     |

#### L40S

**Topology `4×4`** (4 VLM + 4 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 15    | 36    | 41     |

**Topology `2×2`** (2 VLM + 2 LLM GPUs)

| Video length           | 1 min | 5 min | 10 min |
| ---------------------- | ----- | ----- | ------ |
| Target avg latency (s) | 60    | 300   | 600    |
| Max concurrency        | 8     | 17    | 20     |

> **Note**
>
> All measurements use Cosmos3 Nano Reasoner FP8 with Nemotron 3 Nano, 20 frames per 10-second chunk, 9k vision
> tokens, and `max_tokens=100`.