Performance

DeepStream application is benchmarked across various NVIDIA TAO Toolkit and open source models. The measured performance represents end-to-end performance of the entire video analytic application considering video capture and decode, pre-processing, batching, inference, and post-processing to generate metadata. The output rendering is turned off to achieve peak inference performance. For information on disabling the output rendering, see DeepStream Reference Application - deepstream-app chapter.

TAO Pre-trained models

TAO toolkit has a set of pretrained models listed in the table below. If the models below satisfy your requirement, you should start with one of them. These could be used for various applications in smart city or smart places. If your application is beyond the scope of these models, you may re-train one of the popular model architecture using TAO toolkit. The table below shows the end-to-end performance on highly accurate pre-trained models from TAO toolkit. All models are available on NGC. These models are natively integrated with DeepStream and the instructions to run these models are in /opt/nvidia/deepstream/deepstream/samples/configs/tao_pretrained_models/. The following numbers are obtained with sample_1080p_h265.mp4.

Note

The RT-DETR (warehouse) model used with the MV3DT Tracker is a warehouse-centric variant that differs from the RT-DETR model used in the other settings. It is trained for warehouse scenes with classes such as person, forklift and humanoid robots. The model is available on NGC at RT-DETR 2D Warehouse.

Jetson - pretrained models

				Jetson AGX Thor	Jetson AGX Orin
Model Arch	Inference resolution	Precision	Tracker	GPU (FPS)	GPU (FPS)
C-RADIO-B	3*224*224	FP16	No Tracker	1320	544
C-RADIO-L	3*224*224	FP16	No Tracker	557	179
NV-DinoV2-L	3*224*224	FP16	No Tracker	451	126
RT-DETR	3*640*640	FP16	No Tracker	208	96
RT-DETR	3*640*640	FP16	NvDCF Tracker	189	96
RT-DETR (warehouse)	3*544*960	FP16	MV3DT Tracker	95	39
Peoplenet 2.6.3	3*640*640	FP16	MV3DT Tracker	349	147
PeopleNet Transformer	3*544*960	FP16	MV3DT Tracker	25	16
Grounding-DINO	3*544*960	FP16	No Tracker	24	13
MaskGroundingDINOv2	3*544*960	FP16	No Tracker	24	15
TrafficCamnet Transformer Lite	3*544*960	FP16	NvDCF Tracker	153	105
SegFormer	3*640*640	FP16	No Tracker	256	145
Mask2Former + SWIN	3*800*800	FP16	No Tracker	26	18

Performance dgx-spark - pretrained models

				DGX Spark
Model Arch	Inference resolution	Precision	Tracker	GPU (FPS)
C-RADIO-B	3*224*224	FP16	No Tracker	1153
C-RADIO-L	3*224*224	FP16	No Tracker	414
NV-DinoV2-L	3*224*224	FP16	No Tracker	339
RT-DETR	3*640*640	FP16	No Tracker	192
RT-DETR	3*640*640	FP16	NvDCF Tracker	190
RT-DETR (warehouse)	3*544*960	FP16	MV3DT Tracker	106
Peoplenet 2.6.3	3*640*640	FP16	MV3DT Tracker	341
PeopleNet Transformer	3*544*960	FP16	MV3DT Tracker	26
Grounding-DINO	3*544*960	FP16	No Tracker	24
MaskGroundingDINOv2	3*544*960	FP16	No Tracker	25
TrafficCamnet Transformer Lite	3*544*960	FP16	NvDCF	144
SegFormer	3*640*640	FP16	No Tracker	226
Mask2Former + SWIN	3*800*800	FP16	No Tracker	29

Performance dgpu- pretrained models

					RTX 4500	PRO 6000 WS	PRO 6000 SE	L40s	B200	GB200
Model Arch	Inference resolution	Precision	Inference Engine	Tracker	GPU (FPS)	GPU (FPS)	GPU (FPS)	GPU (FPS)	GPU (FPS)	GPU (FPS)
C-RADIO-B	3*224*224	FP16	TRT	No Tracker	2182	4720	4241	2969	10023	11180
C-RADIO-L	3*224*224	FP16	TRT	No Tracker	668	1542	1998	1355	3841	4362
NV-DinoV2-L	3*224*224	FP16	TRT	No Tracker	534	1369	1728	1027	3666	4067
RT-DETR	3*640*640	FP16	TRT	No Tracker	380	1063	1037	643	1448	1541
RT-DETR	3*640*640	FP16	TRT	NvDCF Tracker	379	1024	946	642	1417	1529
RT-DETR (warehouse)	3*544*960	FP16	TRT	MV3DT Tracker	259	660	538	267	955	1513
Peoplenet 2.6.3	3*640*640	FP16	TRT	MV3DT Tracker	741	1899	1524	1114	3572	TBU
Peoplenet Transformer	3*640*640	FP16	TRT	MV3DT Tracker	85	224	205	142	243	217
Grounding-DINO	3*544*960	FP16	TRT	No Tracker	69	178	159	107	219	235
TrafficCamNet Transformer Lite	3*544*960	FP16	TRT	NvDCF Tracker	373	994	920	643	2104	1319
SegFormer	3*640*640	FP16	TRT	No Tracker	560	1278	1291	986	1290	1532
Mask2Former + SWIN	3*800*800	FP16	TRT	No Tracker	68	118	108	75	116	189
MaskGroundingDINO V2	3*544*960	FP16	TRT	No Tracker	69	180	159	107	220	236

Note

TBU (To Be Updated) indicates numbers that are pending an update and will be published in a future revision.

TAO Fine-tuned models

TAO Toolkit Finetuning Microservice provides a new interface for accelerating model training and automating model fine-tuning flows. The fine-tuned models can be used by DeepStream SDK 9.1 out-of-the-box via Inference Builder. The table mentioned in the TAO Pre-trained models section shows the end-to-end performance on fine-tuned models from TAO toolkit.

DeepStream reference model and tracker

DeepStream SDK ships with a reference DetectNet_v2-ResNet10 model and three ResNet18 classifier models. The detailed instructions to run these models with DeepStream are provided in the next section. DeepStream provides four reference trackers: IOU, NvSORT, NvDeepSORT and NvDCF. For more information about trackers, See the Gst-nvtracker section.

Configuration File Settings for Performance Measurement

To achieve peak performance, make sure the devices are properly cooled. For Turing and Ampere GPUs, make sure you use a server that meets the thermal and airflow requirements. Along with the hardware setup, a few other options in the config file need to be set to achieve the published performance. Make the required changes to one of the config files from DeepStream SDK to replicate the peak performance.

Turn off output rendering, OSD, and tiler

OSD (on-screen display) is used to display bounding box, masks, and labels on the screen. If output rendering is disabled, creating bounding boxes is not required unless the output needs to be streamed over RTSP or saved to disk. Tiler is used to display the output in NxM tiled grid. It is not needed if rendering is disabled. Output rendering, OSD and tiler use some percentage of compute resources, so it can reduce the inference performance.

To disable OSD, tiled display and output sink, make the following changes in the DeepStream config file.

• To disable OSD, change enable to 0

  [osd]
  enable=0
  

• To disable tiling, change enable to 0

  [tiled-display]
  enable=0
  

• To turn-off output rendering, change the sink to fakesink.

  [sink0]
  enable=1
  #Type - 1=FakeSink 2=EglSink 3=File
  type=1
  sync=0
  

Use the max_perf setting for tracker

DeepStream SDK 6.2 onwards introduces a new reference low-level tracker library, NvMultiObjectTracker, along with a set of configuration files:

• config_tracker_IOU.yml

• config_tracker_NvDCF_max_perf.yml

• config_tracker_NvDCF_perf.yml

• config_tracker_NvDCF_accuracy.yml

To achieve the peak performance shown in the table above when using the NvDCF tracker, make sure the max_perf configuration is used with video frame resolution matched to that of the inference module. If the inference module uses 480x272 resolution, for example, it would be recommended to use a reduced resolution (e.g., 480x288) for the tracker module like the following:

[tracker]
enable=1
tracker-width=480
tracker-height=288
ll-lib-file=/opt/nvidia/deepstream/deepstream/lib/libnvds_nvmultiobjecttracker.so
#ll-config-file=/opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app/config_tracker_IOU.yml
ll-config-file=/opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app/config_tracker_NvDCF_max_perf.yml
gpu-id=0
enable-batch-process=1
display-tracking-id=1

When the IOU tracker is used, the video frame resolution doesn’t matter, and the default config_tracker_IOU.yml can be used.

CudaDeviceScheduleBlockingSync flag is set by default on dGPU

• On dGPU only, cudaDeviceScheduleBlockingSync flag is set by default on the GPU where the Deepstream pipeline runs. In general, for pipelines with multiple streams, this helps in reducing the CPU utilization without affecting the performance much.

• Setting cudaDeviceScheduleBlockingSync flag when sub batches are enabled in the tracker, results in significant reduction in CPU utilization with similar or negligible dip in performance.

• When the environment variable NVDS_DISABLE_CUDADEV_BLOCKINGSYNC is set to 1, cudaDeviceScheduleBlockingSync flag is not set by default.

• There is a remote possibility that setting cudaDeviceScheduleBlockingSync flag might affect the pipeline performance negatively when the pipeline already runs with GPU utilization close to 100%. Hence, when the user encounters a situation where a Deepstream pipeline is GPU bound and the GPU utilization does not reach close to 100%, then the user may experiment with setting NVDS_DISABLE_CUDADEV_BLOCKINGSYNC to 1 and check if it helps in improving the performance of the pipeline.