CanIEmbody
Matrix/Compute/Edge

What can I run on Jetson AGX Thor 128GB?

128 GB unified memoryBlackwell (2560-core GPU, fifth-gen Tensor Cores)2070 TFLOPS (FP4—Sparse)40 W–130 WView matrix for this target →

Jetson AGX Thor Developer Kit / Jetson T5000 module (128 GB LPDDR5X, 273 GB/s). The smaller Jetson T4000 (64 GB, 1200 FP4-sparse TFLOPS, 40–70 W) is a different target. Headline figure is FP4 with sparsity; BF16/FP16 throughput is far lower. Record power mode (e.g. MAXN) and jetson_clocks state with every run.

Policy by policy

PolicyParamsInference memory (sourced, with precision)On Jetson AGX Thor 128GBBest measured
GR00T N1.7 VLA3B≥16 GBmeasured28.09 Hz · TensorRT · mixed · 2 cam
π0.5 VLA—≥8 GBmeasured7.53 Hz · llama.cpp · fp16 · 2 cam
ACT Imitation80M211 MB measured fp32no data
Diffusion Policy Imitation—1.12 GB measured fp32no data
EO-1 VLA3B≥6.5 GBno data
Evo-1 VLA770M—no data
Fast-WAM World model6B—no data
HIL-SERL RL——no data
LingBot-VA World model5.3B≥18 GBno data
MolmoAct2 VLA—12.1 GB measured bf16no data
Multitask DiT Imitation450M—no data
OpenVLA VLA7.5B15 GB measured bf16no data
OpenVLA-OFT VLA—≥16 GBno data
SmolVLA VLA450M1.75 GB measured fp32no data
TD-MPC RL——no data
VLA-JEPA VLA——no data
VQ-BeT Imitation——no data
WALL-OSS VLA——no data
X-VLA VLA900M—no data
π0 VLA3.3B13.32 GB measured fp32no data
π0-FAST VLA—≥8 GBno data

“Exceeds memory” compares a policy's upstream-stated minimum inference memory (never a measured peak, which is precision-specific) with this target's unified (shared CPU/GPU) memory, so the real budget is lower. It is arithmetic, not observation, and is labelled inferred. Quantised or split runtimes can change the answer; a measured run always wins.

All measured runs on Jetson AGX Thor 128GB

PolicyRuntimePrecisionCamsp50 / p95 msInferencePlatformSource
GR00T N1.7TensorRTmixed2 @256x25635.6 / —28.09 Hzjetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0
conditions

torch_threads: 8 intra-op / 1 inter-op

Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Mixed NVFP4 quantization (enum has no fp4 value).

T3 · community reported
GR00T N1.7TensorRTbf162 @256x25676.8 / —13.02 Hzjetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0
conditions

torch_threads: 8 intra-op / 1 inter-op

Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).

T3 · community reported
GR00T N1.7torch.compilebf162 @256x256100.5 / —9.95 Hzjetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0
conditions

torch_threads: 8 intra-op / 1 inter-op

Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).

T3 · community reported
GR00T N1.7PyTorch eagerbf162 @256x256106.8 / —9.36 Hzjetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0
conditions

torch_threads: 8 intra-op / 1 inter-op

Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).

T3 · community reported
π0.5llama.cppfp162 @224x224132.8 / —7.53 Hz—
conditions

board: Jetson Thor (paper/README: "Jetson Thor")

Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Later configuration from the EagleVLA-Edge README (PR #6); the README notes an accuracy caveat for the two-switch GEMM preset on Thor. Naive PyTorch π0.5 on the same board per the paper: 457.9 ms (2.18 Hz) (not filed as a separate run: the baseline precision is not stated).

T3 · community reported
π0.5llama.cppfp162 @224x224202.8 / —4.93 Hz—
conditions

board: Jetson Thor (paper/README: "Jetson Thor")

Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Paper's 'control frequency' (10.25 Hz) is 1/reaction time with VLM-skipping scheduling; not executed actions/s. Board inferred to be the 128 GB T5000 class from the paper's 130 W max-power entry for Thor (T4000 is 40–70 W). Naive PyTorch π0.5 on the same board per the paper: 457.9 ms (2.18 Hz) (not filed as a separate run: the baseline precision is not stated).

T3 · community reported

Sources

  1. T1 · upstream documentedvendor-docsverified 2026-10-24 (1d ago)

    Jetson Thor series spec table: AGX Thor Developer Kit / T5000 = 2070 TFLOPS (FP4, sparse); T4000 = 1200.

    “Jetson AGX Thor Developer Kit Jetson T5000 Jetson T4000 AI Performance 2070 TFLOPS (FP4—Sparse) 1200 TFLOPS (FP4—Sparse)”
    https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
  2. T1 · upstream documentedvendor-docsverified 2026-10-24 (1d ago)

    Spec table Memory row: 128 GB (Dev Kit / T5000) and 64 GB (T4000) LPDDR5X at 273 GB/s.

    “Memory 128 GB 256-bit LPDDR5X 273 GB/s 64 GB 256-bit LPDDR5X 273 GB/s”
    https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
  3. T1 · upstream documentedvendor-docsverified 2026-10-24 (1d ago)

    Spec table Power row: 40 W–130 W (Dev Kit, T5000), 40 W–70 W (T4000).

    “Power 40 W–130 W 40 W–130 W 40 W–70 W”
    https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/