Isaac Sim Benchmarks#

This page contains key performance indicators (KPIs) for Isaac Sim, captured across different reference hardware and measured using the isaacsim.benchmark.services extension. It also contains a guide on how to collect the same KPIs on your hardware, to compare to our published performance specs.

GPU-Independent KPIs#

These KPIs measure Isaac Sim performance independent of the GPU on which Isaac Sim is running.

Note

These KPIs were measured on a standardized reference machine using an Intel i9-14900k CPU and 32GB of DDR5 RAM.

GPU-Independent KPIs#

Name

Definition

Units

Value

Binary package size (Windows)

Size of Windows binary package

GB

6.47

Binary package size (Ubuntu)

Size of Ubuntu binary package

GB

6.85

Docker container size

Size of Docker container before extraction on NGC

GB

8.54

pip package size

Size of pip package as downloaded

GB

5.06

Startup time (async)

Time from launching Isaac Sim executable to app ready appearing in logs

seconds

33.701 [1]
5.73 [2]

Startup time (non-async)

Time from initializing SimulationApp in standalone Python to app ready appearing in logs

seconds

278 [3]
3.96 [4]

GPU-Dependent KPIs#

These KPIs measure Isaac Sim performance on reference hardware, including frame rate for benchmark scenes and render rate for specific sensor combinations. KPIs are reported as the average KPI value across 600 frames.

Note

SDG KPIs are measured using two (2) 720p cameras in a blank scene with 500 objects generated each frame.

Note

ROS2 Render & Publishing Speed is measured using the reference Nova Carter asset, rendering all available cameras and lidars and publishing all possible messages.

Note

These KPIs were measured on a standardized reference machine using an Intel i9-14900k CPU and 32GB of DDR5 RAM.

Hardware-Dependent KPIs#

Name

Definition

Units

Windows

Ubuntu

Full Warehouse Sample Scene Load Time

Wall-clock time to load Full Warehouse Sample Scene

Seconds

110.9

575.5

Full Warehouse Sample Scene FPS

Frame rate of Full Warehouse Sample Scene

Frames per second

88.7

85.9

Physics steps per second

Number of physics steps executed per wall-clock second in O3dyn sample scene

Hz

578.0

621.1

Isaac ROS Sample Scene FPS

Frame rate of Isaac ROS Sample Scene

Frames per second

14.4

14.2

ROS2 render & publishing speed

Number of pixels rendered and published via ROS2 bridge from Nova Carter ROS asset, per wall-clock second

Megapixels per second

29.7

N/A [5]

SDG images per second (simple)

Number of pixels rendered by SDG per second, with only RGBD annotators enabled, per wall-clock second

Megapixels per second

44.1

43.1

SDG images per second (complex)

Number of pixels rendered by SDG per second, with all annotators enabled, per wall-clock second

Megapixels per second

2.0

2.7

Note

These KPIs were measured on a standardized reference machine using an Intel i9-14900k CPU and 32GB of DDR5 RAM.

Hardware-Dependent KPIs#

Name

Definition

Units

Windows

Ubuntu

Full Warehouse Sample Scene Load Time

Wall-clock time to load Full Warehouse Sample Scene

Seconds

88.7

70.1

Full Warehouse Sample Scene FPS

Frame rate of Full Warehouse Sample Scene

Frames per second

113.6

119.1

Physics steps per second

Number of physics steps executed per wall-clock second in O3dyn sample scene

Hz

571.4

636.9

Isaac ROS Sample Scene FPS

Frame rate of Isaac ROS Sample Scene

Frames per second

27.7

27.4

ROS2 render & publishing speed

Number of pixels rendered and published via ROS2 bridge from Nova Carter ROS asset, per wall-clock second

Megapixels per second

67.5

65.8

SDG images per second (simple)

Number of pixels rendered by SDG per second, with only RGBD annotators enabled, per wall-clock second

Megapixels per second

88.7

88.3

SDG images per second (complex)

Number of pixels rendered by SDG per second, with all annotators enabled, per wall-clock second

Megapixels per second

1.9

2.9

Note

These KPIs were measured on a standardized reference machine using an Intel i9-14900k CPU and 32GB of DDR5 RAM.

Hardware-Dependent KPIs#

Name

Definition

Units

Windows

Ubuntu

Full Warehouse Sample Scene Load Time

Wall-clock time to load Full Warehouse Sample Scene

Seconds

97.5

57.9

Full Warehouse Sample Scene FPS

Frame rate of Full Warehouse Sample Scene

Frames per second

113.1

119.1

Physics steps per second

Number of physics steps executed per wall-clock second in O3dyn sample scene

Hz

581.4

617.3

Isaac ROS Sample Scene FPS

Frame rate of Isaac ROS Sample Scene

Frames per second

37.7

37.8

ROS2 render & publishing speed

Number of pixels rendered and published via ROS2 bridge from Nova Carter ROS asset, per wall-clock second

Megapixels per second

108.6

101.6

SDG images per second (simple)

Number of pixels rendered by SDG per second, with only RGBD annotators enabled, per wall-clock second

Megapixels per second

112.9

115.5

SDG images per second (complex)

Number of pixels rendered by SDG per second, with all annotators enabled, per wall-clock second

Megapixels per second

1.9

2.8

Note

These KPIs were measured on a server using a Xeon W9-3495X CPU with 128GB of DDR5 RAM, on Ubuntu 22.04. Some KPIs are measured on multi-GPU configurations, typically for 2 or 4 GPUs.

Hardware-Dependent KPIs by GPU Count#

Name

Definition

Units

x1

x2

x4

Full Warehouse Sample Scene Load Time

Wall-clock time to load Full Warehouse Sample Scene

Seconds

251.4

254.1

355.5

Full Warehouse Sample Scene FPS

Frame rate of Full Warehouse Sample Scene

Frames per second

110.4

111.5

108.7

Physics steps per second

Number of physics steps executed per wall-clock second in O3dyn sample scene

Hz

321.5

317.5

320.5

Isaac ROS Sample Scene FPS

Frame rate of Isaac ROS Sample Scene

Frames per second

28.0

32.2

35.5

ROS2 render & publishing speed

Number of pixels rendered and published via ROS2 bridge from Nova Carter ROS asset, per wall-clock second

Megapixels per second

80.0

139.9

226.2

SDG images per second (simple)

Number of pixels rendered by SDG per second, with only RGBD annotators enabled, per wall-clock second

Megapixels per second

50.8

58.7

60.1

SDG images per second (complex)

Number of pixels rendered by SDG per second, with all annotators enabled, per wall-clock second

Megapixels per second

23.7

34.6

36.7

Note

These KPIs were measured on a standardized OVX machine using 2x Intel 8362 CPU and 1024GB of DDR4 RAM, on Ubuntu 22.04. Some KPIs are measured on multi-GPU configurations, typically for 2, 4, or 8 GPUs.

Hardware-Dependent KPIs by GPU Count#

Name

Definition

Units

x1

x2

x4

x8

Full Warehouse Sample Scene Load Time

Wall-clock time to load Full Warehouse Sample Scene

Seconds

128.2

128.8

125.1

127.5

Full Warehouse Sample Scene FPS

Frame rate of Full Warehouse Sample Scene

Frames per second

47.4

40.2

45.8

46.6

Physics steps per second

Number of physics steps executed per wall-clock second in O3dyn sample scene

Hz

135.0

125.9

120.8

117.9

Isaac ROS Sample Scene FPS

Frame rate of Isaac ROS Sample Scene

Frames per second

17.3

16.6

19.8

17.3

ROS2 render & publishing speed

Number of pixels rendered and published via ROS2 bridge from Nova Carter ROS asset, per wall-clock second

Megapixels per second

82.9

181.6

214.4

213.6

SDG images per second (simple)

Number of pixels rendered by SDG per second, with only RGBD annotators enabled, per wall-clock second

Megapixels per second

23.7

34.6

36.7

32.0

SDG images per second (complex)

Number of pixels rendered by SDG per second, with all annotators enabled, per wall-clock second

Megapixels per second

0.69

0.71

0.69

0.69

Measuring KPIs on Local Hardware#

Isaac Sim KPIs can be measured using the Python scripts provided in standalone_examples/benchmarks. The KPIs provided above are computed using the scripts as follows. Commands are providing in bash syntax (for Ubuntu); for Windows, replace .sh with .bat and \ for multiline commands to `.

  • Startup time (async): Measured as Runtime for phase: startup from the logs after running

    ./isaac-sim.sh --no-window --/app/quitAfter=200 --/app/file/ignoreUnsavedOnExit=1 \
  --enable isaacsim.benchmark.service

  • Startup time (non-async): Measured as Runtime for phase: startup from the logs after running

    ./python.sh standalone_examples/api/isaacsim.simulation_app/hello_world.py \
  --enable isaacsim.benchmark.services

  • Full Warehouse Sample Scene load time & FPS: Load time is measured as Runtime for phase: loading and FPS is measured as Mean FPS for phase: benchmark.

    ./python.sh standalone_examples/benchmarks/benchmark_scene_loading.py \
  --env-url /Isaac/Environments/Simple_Warehouse/full_warehouse.usd

  • Physics Steps per second: Physics steps per second is measured as 1000.0/Mean Physics Frametime for phase: benchmark.

    ./python.sh standalone_examples/benchmarks/benchmark_robots_o3dyn.py \
  --num-robots 2 --num-gpus 1

  • Isaac ROS Sample Scene FPS: FPS is measured as Mean FPS for phase: benchmark.

    ./python.sh standalone_examples/benchmarks/benchmark_scene_loading.py \
  --env-url /Isaac/Samples/ROS2/Scenario/carter_warehouse_apriltags_worker.usd

  • ROS2 Render & Publishing Speed: First, measure FPS as Mean FPS for phase: benchmark. Then, compute this metric as [FPS] * 8 * 1920 * 1200 / 1e6, accounting for rendering 4 Hawk stereo camera pairs.

    ./python.sh standalone_examples/benchmarks/benchmark_robots_nova_carter_ros2.py \
  --num-robots 1 --enable-3d-lidar 1 --enable-2d-lidar 2 --enable-hawks 4

  • SDG images per second (simple): First, measure FPS as Mean FPS for phase: benchmark. Then compute this metrics as [FPS] * 2 * 1280 * 720 / 1e6, accounting for rendering 2 720p cameras.

    ./python.sh standalone_examples/benchmarks/benchmark_sdg.py \
  --num-cameras 2 --resolution 1280 720 --asset-count 100 \
  --annotators rgb distance_to_image_plane  --disable-viewport-rendering \
  --skip-write --headless

  • SDG images per second (complex): First, measure FPS as Mean FPS for phase: benchmark. Then compute this metrics as [FPS] * 2 * 1280 * 720 / 1e6, accounting for rendering 2 720p cameras.

    ./python.sh standalone_examples/benchmarks/benchmark_sdg.py \
  --num-cameras 2 --resolution 1280 720 --asset-count 100 \
  --annotators all  --disable-viewport-rendering \
  --skip-write --headless
On this page