Hardware Recommendations

August 20, 2026 · View on GitHub

GR00T N1.7 has two hardware profiles: fine-tuning (needs GPU VRAM and compute) and inference/deployment (needs low latency). This guide helps you choose the right hardware for each.

Workflow Diagram

Inference Hardware

Minimum: 1 GPU with 16 GB+ VRAM, CUDA 12.6+.

The table below summarizes end-to-end inference frequency across tested platforms (GR00T N1.7, 4 denoising steps, 1 camera):

Platform	VRAM	PyTorch Eager	With TensorRT	Use Case
H100 80GB HBM3	80 GB	11.7 Hz	35.9 Hz	High-frequency control, multi-env batch inference
H20 96GB HBM3	96 GB	12.0 Hz	29.4 Hz	Cost-effective datacenter inference
RTX Pro 6000 Blackwell	96 GB	12.8 Hz	35.9 Hz	Workstation inference, development
RTX Pro 5000 72GB	72 GB	7.9 Hz	24.7 Hz	Workstation inference
L40	48 GB	7.8 Hz	26.0 Hz	Cloud inference
L20	48 GB	7.1 Hz	23.3 Hz	Cloud inference
DGX Spark	128 GB shared	7.9 Hz	10.1 Hz	Desktop edge, prototyping
AGX Thor	128 GB shared	8.9 Hz	12.4 Hz	Robot-mounted edge deployment
Orin	64 GB shared	2.9 Hz	6.6 Hz	Legacy Jetson edge

Which frequency is this? The rates above are the model inference (replanning) rate — how often the policy produces a new action chunk. This is not the same as the robot's action-execution rate or camera-capture rate (both ~30 FPS in the real-world deployment guide). Because each inference returns a multi-step action chunk, a ~10 Hz inference rate can sustain ~30 FPS execution via action chunking + asynchronous inference — you do not need 30 Hz inference to execute at 30 FPS.

Key Insights

30+ Hz — high-frequency (H100, RTX Pro 6000 with TensorRT): headroom for reactive, low-latency closed-loop control where sub-30 ms per-step latency matters.
10+ Hz — recommended minimum (Thor, Spark with TRT; most dGPUs with torch.compile): sufficient inference rate for typical manipulation tasks (paired with action chunking to reach ~30 FPS execution).
< 10 Hz (Orin with TensorRT): only suitable for slow, non-reactive tasks.
TensorRT Full Pipeline provides 1.5--3.3x speedup over PyTorch Eager depending on platform. Biggest gains are on datacenter GPUs where backbone acceleration is significant.
torch.compile is a good zero-effort middle ground (no engine build step), achieving 1.1--1.9x speedup across all platforms.

For full per-component latency breakdown, see the Deployment Benchmark Results.

Fine-Tuning Hardware

Minimum: 1 GPU with 40 GB+ VRAM. GR00T N1.7 is a ~3B parameter model (bfloat16).

Setup	GPUs	VRAM per GPU	Global Batch Size	Notes
Quick start / prototyping	1x H100, L40, or A100	40--80 GB	32	Single GPU; sufficient for demo datasets
Recommended	4--8x H100 or L40	40--80 GB each	64--640	Multi-GPU via torchrun; faster convergence
Full scale	8x RTX Pro 6000 or DGX	96 GB each	640	Large datasets, production fine-tuning

Key Details

Default fine-tuning tunes the projector + diffusion action head (not the full LLM backbone), keeping peak VRAM under ~35 GB per GPU.
Enabling --tune-llm or --tune-visual significantly increases VRAM — 80 GB+ per GPU recommended.
--gradient-accumulation-steps can compensate for fewer GPUs. For example, 4 GPUs with 8 accumulation steps and per-GPU batch of 8 gives an effective global batch size of 256.
Reduce --num-shards-per-epoch if host memory (not VRAM) is limited — this controls how much dataset is preloaded into RAM.

Software Requirements

Requirement	Version
Python	3.12
CUDA	12.6+ (dGPU) / 13.2 (Thor, Orin) / 13.0 (Spark)
PyTorch	2.7+
OS	Ubuntu 22.04+ (dGPU), JetPack 7.2 (Orin), Ubuntu 24.04 (Thor, Spark)
Package manager	uv (recommended)

Platform-specific installation instructions: see the Deployment Guide.

Recommended Configurations

Starter Kit

For development, small-scale fine-tuning, and edge deployment:

Component	Recommendation
Training	1--4x L40 (48 GB) or RTX Pro 5000/6000 workstation
Edge Deployment	Jetson AGX Thor Developer Kit (128 GB shared memory, Blackwell GPU)
Storage	500 GB+ SSD (datasets + checkpoints)

Center of Excellence

For production fine-tuning and high-throughput inference:

Component	Recommendation
Training	DGX with 8x H100/B200, or RTX Pro Server with 8x RTX Pro 6000 Blackwell
Inference Server	H100 or H20 node with TensorRT Full Pipeline (35+ Hz per GPU)
Edge Deployment	Jetson AGX Thor or DGX Spark
Storage	Scalable networked storage (NFS/S3) for large-scale datasets