Jetson AGX Orin 64GB for Edge AI — Specs, Sizing & Deployment Fit
Last updated: August 2026
The high-density workhorse: 275 TOPS, 64 GB of unified memory, and ~20-stream planning capacity for heavy multi-model pipelines and ≤13B-parameter LLMs.
Planning Takeaway
The high-density workhorse: 275 TOPS, 64 GB of unified memory, and ~20-stream planning capacity for heavy multi-model pipelines and ≤13B-parameter LLMs. Since July 2026 it shares its ~$2,999 module price with the newer Jetson T4000 — a real decision point for new designs.
Specifications
| Spec | Value |
|---|---|
| Compute | 275 INT8 TOPS (Ampere GPU + 2× NVDLA) |
| Memory | 64 GB LPDDR5, ~204.8 GB/s |
| Power modes | 15W · 30W · 40W · 50W · MAXN (15–60 W envelope) |
| Video decode (modeled planning ceiling) | ~13× 1080p / ~5× 4K H.264 concurrent |
| Est. planning capacity | ~20 concurrent 1080p detection streams |
| Cooling | Active required |
| Form factor | Jetson AGX module (100×87 mm) |
| Runtimes / precisions | TensorRT, PyTorch, ONNX, llama.cpp · FP32/FP16/INT8 |
| Software | JetPack 7.2 (Ubuntu 24.04, CUDA 13, June 2026) — current; JetPack 6.x (Ubuntu 22.04) — legacy track. Sibling AGX Orin 32GB module (~$1,799) gained MAXN_SUPER on JetPack 7.2 (200 → 241 TOPS) |
| Indicative price | ~$2,999 module / ~$3,499 dev kit (as of mid-2026) |
Sizing Fit
Suits 16-camera-class analytics nodes, multi-model + VLM co-location within 64 GB, and head-node roles in hierarchical edge deployments.
- 16-camera smart-city traffic architecture — Reference architecture
- Thor vs AGX Orin decision framework — Comparison
- Which VLM fits which Jetson — Guide
- Multi-model inference on Jetson — Guide
Decision Framework
Choose the AGX Orin 64GB if:
- You need 16+ streams or heavy multi-model pipelines with headroom on one node.
- Your workload includes quantized LLMs up to roughly the 13B–30B (GGUF Q4) class alongside vision.
- You want the mature, widely deployed AGX ecosystem, carrier options, and a dev kit (~$3,499).
Choose a sibling instead if:
- Jetson T4000 — starting a new design at the same ~$2,999: newer Blackwell-generation FP4 compute in a lower power envelope, at the cost of a younger ecosystem.
- Jetson AGX Thor T5000 — you need 70B-class on-device LLMs, VLA robotics models, or 128 GB memory.
- Jetson Orin NX 16GB — your node serves ≤12 streams — the ~$999 NX is the better-balanced spend.
Frequently Asked Questions
How much does the Jetson AGX Orin 64GB cost?
As of the July 2026 NVIDIA repricing: ~$2,999 for the module (up from $1,599) and ~$3,499 for the developer kit. The AGX Orin 32GB module lists at ~$1,799.
AGX Orin 64GB or Jetson T4000 — they cost the same?
Both modules list at ~$2,999 as of mid-2026. The AGX Orin 64GB offers the mature ecosystem, 64 GB, and proven carrier/thermal designs; the T4000 offers newer-generation compute (up to ~1200 FP4 TFLOPs) in a lower power envelope but a younger deployment ecosystem. For established vision fleets choose AGX; for new long-horizon designs weigh the T4000 seriously.
Can the AGX Orin 64GB run large language models?
Yes, within limits: 13B-class models run comfortably quantized, and ~30B-class fits with GGUF Q4. It lacks FP4 hardware, so 70B-class on-device models are Thor territory.
What JetPack version does the AGX Orin run?
JetPack 7.2 (June 2026) is current for this module — Ubuntu 24.04, Linux kernel 6.8, CUDA 13. JetPack 6.x remains a widely deployed legacy track on Ubuntu 22.04. Moving an existing JetPack 6 install to 7 requires the OTA package or a full reflash — a plain apt upgrade is not an upgrade path. JetPack 7.2 also added MAXN_SUPER on the sibling AGX Orin 32GB, lifting it from 200 to 241 TOPS.