Recorded comparisons

Faster inference in action.

Compare baseline and ZETIC-optimized execution on the same hardware.

11 demos

π0.5

VLA

Jetson AGX Thor

9.7× faster

vs PyTorch eager

253.7 ms 26.08 ms

Policy inference · p50

LinkedIn X Email

LingBot-VA

VLA

Jetson AGX Thor

7.2× faster

vs PyTorch reference

11.15 s 1.54 s

Inference per action chunk

LinkedIn X Email

Sonata (PTv3)

Point-cloud segmentation

Jetson AGX Thor

23.1× faster

vs PyTorch FP32 · CPU voxelization

715 ms 31 ms

End-to-end latency per frame

LinkedIn X Email

Isaac-0.2-2B + SAM2

Vision & tracking

Jetson Thor

10.5× faster · peak

vs TensorRT FP16 · full generation

3.16 s 0.30 s

Prompt response

LinkedIn X Email

Cosmos Policy

VLA

Jetson Thor

6.7× faster

vs As-released · skip-decode

1.83 s 275 ms

Inference per 16-action chunk · p50

LinkedIn X Email

FoundationStereo

Stereo depth

Jetson AGX Thor

2.7× faster

vs PyTorch FP16

104 ms 38 ms

Latency per frame

LinkedIn X Email

FoundationStereo · Three modes

Stereo depth

Jetson AGX Orin

3.6–8.6× faster

vs PyTorch reference

320 ms 37–89 ms

Inference latency per frame

LinkedIn X Email

π0.5 · Orin

VLA

Jetson AGX Orin 64GB

2.8× faster

vs TensorRT FP16

215 ms 77 ms

Policy inference latency

LinkedIn X Email

Sonata (PTv3) · Orin

Point-cloud segmentation

Jetson AGX Orin

6.7× faster

vs FP32 · CPU voxelization

991 ms 149 ms

End-to-end latency per frame

LinkedIn X Email

YOLO26m · 500 images

Object detection

Jetson AGX Orin

1.6× faster

vs TensorRT INT8 · detection head FP16

4.83 ms 2.99 ms

Model latency per image

LinkedIn X Email

YOLO26m · 10-camera video

Object detection

Jetson AGX Orin

1.6× faster

vs TensorRT INT8 · detection head FP16

207 fps 335 fps

Model inference capacity · frames/s

LinkedIn X Email

Looking for Orin, Nano, or precision-specific benchmarks?

Explore ixian-bench ↗

VLA

π0.5

π0.5 · Jetson AGX Thor · PyTorch eager baseline

9.7×faster

vs PyTorch eager

253.7 ms → 26.08 ms

Policy inference · p50

Measurement details

Model & hardware: π0.5 · Jetson AGX Thor. Model by Physical Intelligence.
Execution: PyTorch eager → ZETIC hardware-aware execution optimization.
Measurement: Policy inference latency (p50): 253.7 ms → 26.08 ms. These values measure inference, not complete physical task duration.
Recording: Simulated robot actions replayed with each side’s measured inference delay; this is not a live robot recording.
Playback: 1× until the optimized side completes the first task, then 5× and 10×, as labeled in the recording.

Open demo page ↗

LinkedIn X Email

VLA

LingBot-VA

LingBot-VA · Jetson AGX Thor · PyTorch reference baseline

7.2×faster

vs PyTorch reference

11.15 s → 1.54 s

Inference per action chunk

Measurement details

Model & hardware: LingBot-VA · Jetson AGX Thor. Model by Robbyant.
Execution: PyTorch reference → ZETIC optimized runtime.
Measurement: Inference per action chunk: 11.15 s → 1.54 s.
Recording: Simulated robot-task replay under matching conditions. Pauses reflect measured inference time.
Playback: 2× playback, as labeled.

Open demo page ↗

LinkedIn X Email

Point-cloud segmentation

Sonata (PTv3)

Sonata (PTv3) · Jetson AGX Thor · Stock PyTorch baseline

23.1×faster

vs PyTorch FP32 · CPU voxelization

715 ms → 31 ms

End-to-end latency per frame

Measurement details

Model & hardware: Sonata (PTv3) · Jetson AGX Thor · Point-cloud segmentation.
Execution: Stock PyTorch FP32 with CPU voxelization → ZETIC hardware-aware execution optimization.
Measurement: End-to-end latency per frame: 715 ms → 31 ms (23.1×). This includes the frame-processing pipeline, not just model inference.
Recording: 10 Hz sensor stream. Each side shows its latest completed result; frame counters and lag relative to the sensor are shown in the recording.
Playback: 1× playback, as labeled in the recording.

Open demo page ↗

LinkedIn X Email

Vision & tracking

Isaac-0.2-2B + SAM2

Isaac-0.2-2B + SAM2 · Jetson Thor · FP16 full-generation baseline

10.5×faster · peak

vs TensorRT FP16 · full generation

3.16 s → 0.30 s

Prompt response

Measurement details

Model & hardware: Isaac-0.2-2B + SAM2 · NVIDIA Jetson Thor · TensorRT backend.
Execution: Baseline: FP16 response generation completes before coordinate output. Optimized: FP8 response generation stops at coordinates; image understanding remains FP16.
Measurement: Instruction submission to completed target-coordinate output: 3.16 s → 0.30 s. This is prompt-response time, not time to first token.
Recording: The recording shows target selection, segmentation, and tracking following a text instruction.
Playback: 1× real-time playback.

Open demo page ↗

LinkedIn X Email

VLA

Cosmos Policy

Cosmos Policy · Jetson Thor · As-released, skip-decode baseline

6.7×faster

vs As-released · skip-decode

1.83 s → 275 ms

Inference per 16-action chunk · p50

Measurement details

Model & hardware: Cosmos Policy · Jetson Thor. Model by NVIDIA.
Execution: As-released Cosmos Policy with its official skip-decode flag → ZETIC optimized model and runtime. The baseline plans synchronously; the optimized run plans asynchronously.
Measurement: p50 inference latency per 16-action chunk: 1.83 s → 275 ms. Task-completion differences also include the change from synchronous to asynchronous planning.
Recording: Simulation of the same 10 tasks with matching model, starting conditions, and 20 Hz control.
Playback: 1× opening, then 5× and 10× on both sides, as labeled.

Open demo page ↗

LinkedIn X Email

Stereo depth

FoundationStereo

C-Fast-FoundationStereo · Jetson AGX Thor · Stock PyTorch baseline

2.7×faster

vs PyTorch FP16

104 ms → 38 ms

Latency per frame

Measurement details

Model & hardware: C-Fast-FoundationStereo · Jetson AGX Thor.
Execution: Stock PyTorch FP16 → ZETIC hardware-aware execution optimization.
Measurement: Latency per frame: 104 ms → 38 ms (2.7×).
Recording: KITTI stereo frames replayed at 15 Hz, recorded at 10 Hz. Each side shows its latest completed depth map. The frame counters show skipped frames and lag relative to the camera stream.
Playback: 1× playback, as labeled in the recording.

Open demo page ↗

LinkedIn X Email

Stereo depth

FoundationStereo · Three modes

C-Fast-FoundationStereo · Jetson AGX Orin · PyTorch baseline

3.6–8.6×faster

vs PyTorch reference

320 ms → 37–89 ms

Inference latency per frame

Measurement details

Model & hardware: C-Fast-FoundationStereo · NVIDIA Jetson AGX Orin. Model by NVIDIA.
Execution modes: Accurate: 89 ms (3.6×). Balanced: 73 ms (4.4×). Fast: 37 ms (8.6×). PyTorch baseline: 320 ms per frame.
Accuracy evaluation: BP-1 error on Middlebury / ETH3D: Accurate 4.91% / 0.94%; Balanced 5.11% / 1.13%; Fast 8.31% / 1.63%. Lower is better.
Evaluation scope: Separate evaluation: Middlebury 15 scenes / ETH3D 27 pairs. BP-1 counts pixels with disparity error greater than 1 px.
Playback: 1× playback. Speedups compare inference latency, not video playback.

Open demo page ↗

LinkedIn X Email

VLA

π0.5 · Orin

π0.5 · Jetson AGX Orin 64GB · TensorRT FP16 baseline

2.8×faster

vs TensorRT FP16

215 ms → 77 ms

Policy inference latency

Measurement details

Model & hardware: π0.5 · Jetson AGX Orin 64GB. Model by Physical Intelligence.
Execution: TensorRT FP16 baseline → ZETIC hardware-aware execution optimization.
Measurement: Policy inference latency shown in the recording: 215 ms → 77 ms (2.8×). This measures inference, not complete task duration.
Recording: LIBERO-Spatial simulation.
Playback: Playback speed changes as labeled in the video; the speedup compares inference latency, not playback speed.

Open demo page ↗

LinkedIn X Email

Point-cloud segmentation

Sonata (PTv3) · Orin

Sonata (PTv3) · Jetson AGX Orin · FP32 baseline with CPU voxelization

6.7×faster

vs FP32 · CPU voxelization

991 ms → 149 ms

End-to-end latency per frame

Measurement details

Model & hardware: Sonata (PTv3) · Jetson AGX Orin.
Execution: FP32 baseline with CPU voxelization → ZETIC optimized BF16 execution with GPU-resident input.
Measurement: End-to-end latency per frame shown in the recording: 991 ms → 149 ms (6.7×). This includes frame processing, not just model inference.
Recording: Measured replay of a 10 Hz sensor stream. Each side shows its latest completed point-cloud segmentation result.
Playback: 1× playback, as labeled in the recording.

Open demo page ↗

LinkedIn X Email

Object detection

YOLO26m · 500 images

YOLO26m · Jetson AGX Orin · TensorRT INT8 baseline, detection head FP16

1.6×faster

vs TensorRT INT8 · detection head FP16

4.83 ms → 2.99 ms

Model latency per image

Measurement details

Model: YOLO26m by Ultralytics.
Measurement: Model inference time per image, excluding the rest of the image-processing pipeline.
Test setup: The same 500 COCO val2017 images on each side, processed one at a time (batch size 1).

Open demo page ↗

LinkedIn X Email

Object detection

YOLO26m · 10-camera video

YOLO26m · Jetson AGX Orin · TensorRT INT8 baseline, detection head FP16

1.6×faster

vs TensorRT INT8 · detection head FP16

207 fps → 335 fps

Model inference capacity · frames/s

Measurement details

Model: YOLO26m by Ultralytics.
Video scenario: Ten 30 fps camera streams on one Jetson AGX Orin, requiring 300 frames/s in total.
Measurement: 207 → 335 fps of model inference capacity, based on 4.84 → 2.99 ms per frame. This excludes the rest of the video-processing pipeline.
Playback: Playback speed varies as labeled in the recording.

Open demo page ↗

LinkedIn X Email

Faster inference in action.

π0.5

LingBot-VA

Sonata (PTv3)

Isaac-0.2-2B + SAM2

Cosmos Policy

FoundationStereo

FoundationStereo · Three modes

π0.5 · Orin

Sonata (PTv3) · Orin

YOLO26m · 500 images

YOLO26m · 10-camera video

Get faster inference.

Website analytics