> 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. # Alert Verification Performance ## Overview The VSS Alert Bridge verification pipeline processes live RTSP streams through CV detection and analytics (RTVI-CV and video-analytics microservices) and a VLM stage to classify and confirm security alerts. Benchmarks measure end-to-end (E2E) latency, VLM and VST stage latency, and verified alert throughput across the NVIDIA RTX PRO 4500, RTX PRO 6000, DGX H100 SXM, DGX Spark, and AGX Thor platforms using RT-DETR and Grounding DINO (GDINO) with Cosmos3 Nano Reasoner (CR3) precision variants. ## Test Configuration | Parameter | Value | | --------------------- | -------------------------------------------------------------------- | | VSS Release | 3.3 | | Platforms tested | NVIDIA RTX PRO 4500, RTX PRO 6000, DGX H100 SXM, DGX Spark, AGX Thor | | CV models | RT-DETR, Grounding DINO (GDINO) | | VLM | Cosmos3 Nano Reasoner (CR3) | | VLM precision | BF16, FP8, NVFP4 (platform- and model-dependent) | | Stream concurrency | 1–50 concurrent RTSP streams (platform-dependent) | | Configured stream FPS | 10 | | Video resolution | 1920x1080 (1080p) | | OSL | 1 | | Test date | July 2026 | ## Performance by Platform #### RTX PRO 4500 RT-DETR with Cosmos3 Nano Reasoner (CR3). | Streams | Precision | E2E Avg (s) | E2E P90 (s) | E2E P99 (s) | VLM P90 (s) | VST P90 (s) | Verified/min/stream | | ------- | --------- | ----------- | ----------- | ----------- | ----------- | ----------- | ------------------- | | 1 | BF16 | 0.73 | 0.90 | 0.99 | 0.91 | 0.25 | 4.68 | | 40 | BF16 | 12.35 | 25.07 | 36.38 | 9.19 | 0.09 | 2.92 | | 40 | FP8 | 7.97 | 13.46 | 22.27 | 4.73 | 0.09 | 4.58 | | 40 | NVFP4 | 7.28 | 12.33 | 14.74 | 4.71 | 0.09 | 4.60 | #### RTX PRO 6000 RT-DETR with Cosmos3 Nano Reasoner (CR3). | Streams | Precision | E2E Avg (s) | E2E P90 (s) | E2E P99 (s) | VLM P90 (s) | VST P90 (s) | Verified/min/stream | | ------- | --------- | ----------- | ----------- | ----------- | ----------- | ----------- | ------------------- | | 1 | BF16 | 0.51 | 0.90 | 0.99 | 0.45 | 0.26 | 4.48 | | 20 | BF16 | 1.44 | 2.33 | 2.49 | 1.87 | 0.24 | 4.49 | | 30 | BF16 | 1.95 | 3.75 | 4.88 | 2.77 | 0.24 | 4.53 | | 50 | BF16 | 4.62 | 8.97 | 13.08 | 4.55 | 0.17 | 4.73 | | 1 | FP8 | 0.51 | 0.90 | 0.99 | 0.45 | 0.25 | 4.57 | #### DGX H100 SXM RT-DETR with Cosmos3 Nano Reasoner (CR3). | Streams | Precision | E2E Avg (s) | E2E P90 (s) | E2E P99 (s) | VLM P90 (s) | VST P90 (s) | Verified/min/stream | | ------- | --------- | ----------- | ----------- | ----------- | ----------- | ----------- | ------------------- | | 1 | BF16 | 0.58 | 0.90 | 0.99 | 0.45 | 0.28 | 4.32 | | 20 | BF16 | 2.83 | 4.66 | 6.37 | 4.29 | 0.24 | 4.73 | | 30 | BF16 | 4.44 | 8.62 | 9.96 | 7.39 | 0.23 | 4.71 | | 50 | BF16 | 24.63 | 52.56 | 85.41 | 9.12 | 0.10 | 5.52 | | 1 | FP8 | 0.60 | 0.90 | 0.99 | 0.45 | 0.25 | 4.57 | | 20 | FP8 | 2.45 | 4.48 | 4.96 | 3.59 | 0.24 | 4.80 | | 30 | FP8 | 4.14 | 8.22 | 9.83 | 5.10 | 0.23 | 4.81 | | 50 | FP8 | 24.21 | 52.17 | 80.38 | 9.14 | 0.10 | 5.64 | #### DGX Spark GDINO and RT-DETR with Cosmos3 Nano Reasoner (CR3). | CV Model | Streams | Precision | E2E Avg (s) | E2E P90 (s) | E2E P99 (s) | VLM P90 (s) | VST P90 (s) | Verified/min/stream | | -------- | ------- | --------- | ----------- | ----------- | ----------- | ----------- | ----------- | ------------------- | | GDINO | 1 | BF16 | 1.19 | 2.35 | 2.49 | 1.08 | 0.24 | 0.28 | | GDINO | 1 | FP8 | 1.11 | 2.35 | 2.49 | 0.95 | 0.24 | 0.28 | | GDINO | 1 | NVFP4 | 0.99 | 2.12 | 2.46 | 0.95 | 0.24 | 0.28 | | RT-DETR | 1 | BF16 | 0.95 | 0.94 | 2.21 | 0.95 | 0.24 | 0.44 | | RT-DETR | 1 | FP8 | 0.88 | 0.91 | 1.41 | 0.95 | 0.25 | 0.43 | | RT-DETR | 1 | NVFP4 | 0.82 | 0.91 | 1.43 | 0.94 | 0.25 | 0.42 | | RT-DETR | 14 | BF16 | 235.38 | 120.00 | 120.00 | 30.00 | 0.09 | 4.40 | | RT-DETR | 14 | FP8 | 546.79 | 120.00 | 120.00 | 30.00 | 0.09 | 4.05 | | RT-DETR | 14 | NVFP4 | 8.05 | 11.90 | 14.73 | 7.78 | 0.09 | 4.59 | > **Note** > > Where E2E P90 and P99 are identical (for example, 120.00 s at 14 > streams for BF16 and FP8), this is a known limitation of the > bucketing method used for data collection. #### AGX Thor GDINO and RT-DETR with Cosmos3 Nano Reasoner (CR3). | CV Model | Streams | Precision | E2E Avg (s) | E2E P90 (s) | E2E P99 (s) | VLM P90 (s) | VST P90 (s) | Verified/min/stream | | -------- | ------- | --------- | ----------- | ----------- | ----------- | ----------- | ----------- | ------------------- | | GDINO | 1 | BF16 | 3.40 | 4.76 | 4.99 | 1.90 | 0.19 | 0.40 | | GDINO | 1 | FP8 | 3.39 | 4.76 | 4.99 | 1.90 | 0.20 | 0.40 | | GDINO | 1 | NVFP4 | 3.40 | 4.76 | 4.99 | 1.90 | 0.19 | 0.40 | | RT-DETR | 1 | BF16 | 3.39 | 4.76 | 4.98 | 1.90 | 0.24 | 0.67 | | RT-DETR | 1 | FP8 | 3.40 | 4.76 | 4.98 | 1.90 | 0.24 | 0.67 | | RT-DETR | 1 | NVFP4 | 3.42 | 4.76 | 4.99 | 1.90 | 0.24 | 0.66 | > **Note** > > All benchmarks use input streams at 1080p, H.264 input. > > For dGPU platform tests (NVIDIA RTX PRO 4500, RTX PRO 6000, and DGX H100 SXM), > two GPUs were used: one GPU dedicated to RTVI-CV and one GPU dedicated to RTVI-VLM.