Benchmark
This benchmark measures software latency and command timing when whole-body control (WBC) and GR00T N1.7 run together on AGX Thor, comparing shared GPU execution with CUDA Green Context isolation for WBC. Latency is measured from the source callback start to the command sink callback end along two paths:
- WBC (green): joint state through the WBC inference operator to the combined joint command at 50 Hz.
- GR00T image-to-command (teal): image through the GR00T inference operator to its first trace-visible combined joint command, with 5 Hz inference selecting from 30 Hz mock images.
The diagrams highlight each measured path; dotted connections carry validation inputs. Mock inputs and outputs exclude physical camera latency, robot communication, and actuator response.
Results
Jetson AGX Thor Benchmark Configuration
- Run date: October 28–29, 2026 (UTC), one one-hour trial per Green Context condition on AGX Thor.
- Platform: NVIDIA Jetson AGX Thor Developer Kit; NVIDIA driver 595.78;
PREEMPT_RT kernel
6.8.12-1021-rt-tegra. The Jetson Linux package version was not recorded. - Isaac OS commit:
3c3ca720ee5ecd6f47833e44489d7a461ada4030(internal mirror). - Tools: CUDA 13.2, TensorRT 10.16.0.72, Nsight Systems 2026.4.1.
- Container image ID:
sha256:14a7a8df08533ed3e1906b5fe53f564348484f3e32f25252b2b497f8c43606bc. - Holoscan SDK: prebuilt 5.0.0 from
nvcr.io/nvidia/clara-holoscan/holoscan:v5.0.0-cuda13. - Configuration:
config/control.yamlandoperators/gr00t_n1_7/config/n1_7_g1.yamlat the Isaac OS commit above; five GR00T model engines initialized before timed trials. - WBC SM allocation: off: 0 (shared GPU); on: 2 of Thor’s 20 SMs, leaving 18 for the residual graph.
- Host settings: reference 120W power mode (mode 1); CPU 2.601 GHz, GPU GPC 1.386 GHz, GPU NVD 1.557 GHz and EMC 4.266 GHz locked; display manager inactive; CPU deep idle disabled. The shared host preflight passed before and after each capture.
- RT CPU budget:
kernel.sched_rt_runtime_us=-1(RT bandwidth throttling disabled). Timer migration was disabled withkernel.timer_migration=0. - Scheduling: CPUs 8–13 isolated; default IRQ, unbound workqueue and kernel-thread
affinity CPUs 0–7. Controller pinned to CPU 1, background operators to CPU 0,
GR00T to CPU 12 (
SCHED_FIFOpriority 5), and WBC to CPU 13 (SCHED_FIFOpriority 10). Worker affinity and scheduling were verified during each capture. Tickless operation and RCU offload were not enabled. - Measurement: source callback start to sink callback end; a common 20-second warmup and 3,600-second source window for both paths. Each capture stops after 181,100 joint commands.
- GR00T scope: first trace-visible final command for each matched image. Retained action state and fan-in do not establish an image’s contribution to later commands or action playback.
- Measurement order: one trial with Green Context off, then one with it on. All statistics describe the single trial for each condition.
WBC p99.9 latency was 2.5747 ms with Green Context off and 0.3747 ms with Green Context on while GR00T ran concurrently. With Green Context on, WBC p50 was 0.2794 ms; GR00T image-to-command latency was 120.2659 ms at p50 and 140.2846 ms at p99.9. No measured response exceeded the 20 ms WBC or 200 ms GR00T latency budget. Each condition was measured once for one hour.
GR00T commands are forwarded on the 50 Hz WBC update cycle. A result that misses one update waits for the next, which can add about 20 ms to image-to-command latency. In these captures, Green Context reduced WBC tail latency. GR00T image-to-command latency had a lower median but a higher p99.9. These measurements do not establish the cause of differences between machines.
WBC results
GR00T image-to-command results
Full Results
All statistics describe one one-hour trial per condition. Green Context off was measured first, followed by Green Context on; there are no repeated trials.
WBC statistics
GR00T image-to-command statistics
Each trial uses the same one-hour source window after a 20-second warmup. Eligible images arrive at 30 Hz, while GR00T selects images for inference at 5 Hz. Unmatched images have no trace-visible response on this exact path; they do not prove a transport drop. Matched / selected counts refer to first responses. Counts are not pooled across paths.
Run the benchmark
Before benchmarking, check your AGX Thor host for settings that may limit performance. Run this command on the host from the repository root:
Follow Optimize your host for setup guidance. Use the host settings above to reproduce these results.
See Profile an Application for profiling and report options.
Build and capture
Set WBC_SMS=2 for Green Context on, or 0 for off. The unprofiled 500-command run
initializes the models before capture. Verify the SM allocation in the application log.
The 181,100 final joint commands allow a 20-second warmup, one-hour measurement window and about 2 seconds for trailing responses at 50 Hz. Run once with Green Context off and once with it on, using the same build and host settings. Keep a separate output directory for each condition; no repeated trials are needed.
Analyze
From the same host shell, open the capture directory in the app container:
Run all remaining commands in this container shell: export the capture, then generate the WBC and GR00T reports below.
WBC path
Select the joint-state-to-command path with a 20 ms period and latency budget:
GR00T image-to-command path
Select the image-to-first-command path with a 200 ms period and latency budget:
Expect about 180,000 matched WBC messages and 18,000 matched images in each one-hour window. About 108,000 images are eligible at 30 Hz; unmatched images have no visible response on this exact path because inference selects images at 5 Hz.