PyTorch eager
BF16
95.02%191/201
- Backend
- PyTorch eager
- Precision
- BF16
- Tasks
- 10
- Episodes
- 201
VLA · NVIDIA · measured on AGX Thor and AGX Orin
GR00T N1.7 robot foundation model: two camera frames and the robot state in, a decoded action chunk out.
Compare with other models on AGX Thor →
PyTorch eager → ZETIC optimized · AGX Thor
AGX Thor · per policy call
| Metric | ReferencePyTorch eager · BF16 | ZETIC optimizedMixed precision |
|---|---|---|
| Policy latency · p50 | 108.52 ms | 23.46 ms |
| Energy per inference | 7.11 J | 2.13 J |
| Peak memory | 15.57 GB | 10.14 GB |
View optimized build results →
Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.
p50. GR00T N1.7 · LIBERO 10 · AGX Thor
Speedup is reference latency divided by path latency on this board. Lowest recorded latency excludes marked model or inference-schedule variants. Check the configurations below before treating paths as a matched comparison.
PyTorch eager · Model authors · BF16
PyTorch eager · BF16 · batch 1
Remeasurement pending. Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.
ZETIC · ZETIC · Mixed precision
ZETIC · MIXED PRECISION · batch 1
Remeasurement pending. Quality validation pending. Additional validation pending. Peak board memory above idle (MemTotal − MemAvailable), over the full command including model loading; not process RSS.
PyTorch eager: End-to-end latency, p50 (median of round medians), 1 round · 553 samples per round.
ZETIC: End-to-end latency, p50 (median of round medians), 1 round · 2,553 samples per round.
Source: reviewed benchmark measurements. Public measurement: record-b5d142329cc468f9c335b93b6624bb9097f16eaf56323e517756d38cb8d404e3.
27/178 LIBERO-10 fixtures; one instruction (bucket 156).
Measurement methodology ↗GR00T N1.7 · LIBERO 10
AGX Thor
PyTorch eager
BF16
95.02%191/201
ZETIC optimized
Precision not reported
94.55%191/202
Evaluation coverage differs; no paired success-rate delta.
Build with ZETIC
Share your GR00T N1.7 checkpoint for AGX Thor, with goals for lower latency, energy per inference and memory use.