Benchmark
This benchmark measures software latency and command timing in the SO-101 sinus-motion application on AGX Thor, using mock joint states at 100 Hz. Latency is measured from the state source callback start to the command sink callback end. The run instructions select the direct state source → motion planner → command sink path, highlighted below. The sine generator supplies the planner’s latest target; this metric measures the response to incoming state and excludes target generation and target age. Results cover typical and tail latency, message rates, and command-period error; mock inputs and outputs exclude robot communication and physical actuator response.
Results
Jetson AGX Thor Benchmark Configuration
- Run date: October 14, 2026 (UTC), one one-hour trial on AGX Thor.
- Platform: NVIDIA Jetson AGX Thor Developer Kit; Jetson Linux R39.2.1
(
nvidia-l4t-core39.2.1-20260806224157), NVIDIA driver 595.78; PREEMPT_RT kernel6.8.12-1021-rt-tegra. - Isaac OS commit:
fba280c50b89cf2cbbbac3fa62ea470bd5ba2480(internal mirror). - Tools: CUDA 13.2.51, TensorRT 10.16.0.72, Nsight Systems 2026.2.1.210.
- Container image ID:
sha256:72fda6a18147bfa7ada7478141aaf1e619338a19df1dfbbab8c3f617f30788c3. - Holoscan SDK commit:
9be029084c7c0d5fbce67e7eed42d5319f5942dc(5.0.0). - Host settings: 120 W 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.
- RT CPU budget:
kernel.sched_rt_runtime_us=-1(RT bandwidth throttling disabled). - Scheduling: CPUs 8–13 isolated; housekeeping/IRQ CPUs 0–7. No explicit
application CPU affinity. The
nohz_fullandrcu_nocbsboot options are set, but the running kernel lacksCONFIG_NO_HZ_FULLandCONFIG_RCU_NOCB_CPU; tickless operation and RCU offload are not enabled. - Measured path: state source → motion planner → command sink.
- Measurement: one trial, excluding 1,000 warmup messages and measuring the next 360,000 (one hour at 100 Hz). All statistics describe this single trial.
Latency was 0.0271 ms at p50 and 0.0993 ms at p99.9. The one-hour trial contained 360,000 eligible, matched and selected responses, with zero unmatched sources, responses over the 10 ms budget or trace-health warnings.
Full Results
All statistics describe the single one-hour trial. Source counts and latency-budget exceedances refer to the direct state-to-command path.
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 to resolve any warnings before capturing results.
See Profile an Application for profiling and report options.
Build and capture
Capture the full mock graph; analysis selects the highlighted direct state-to-command path.
The 361,100 final commands cover 1,000 warmup messages (10 seconds), 360,000 measured messages (one hour at 100 Hz), and 100 trailing commands. Run one trial with these settings.
Analyze
From the same host shell, open the capture directory in the app container:
Run these commands inside the container. They select the direct state source → planner → sink path, skip 1,000 warmup messages and measure the next 360,000.