armhub.dev
Vendors · 43 pages
Vendors

NVIDIA Jetson — edge AI on Cortex-A

NVIDIA's Jetson family — Cortex-A CPUs plus NVIDIA GPU and DLA on a single SoM. From TK1 in 2014 to Thor in 2025, the JetPack SDK, TensorRT, CUDA, and why Jetson dominates serious edge AI.

5 min read · 1,207 wordsupdated · May 13, 2026

NVIDIA Jetson is the dominant platform for serious edge AI — workloads too large or too latency-sensitive for an MCU-class NPU like MAX78000 or Coral Edge TPU but where shipping data to a cloud GPU is unacceptable. The defining capability is software portability: a model trained on an NVIDIA workstation or H100 cluster deploys to a Jetson with the same CUDA kernels, the same cuDNN, and the same TensorRT inference engine. No cross-vendor conversion, no operator-coverage gaps, no quantisation surprises beyond the ones the developer chooses to introduce.

The hardware platform is a System-on-Module (SoM) — typically 70 x 45 mm for Nano-class and 100 x 87 mm for AGX-class — combining a Cortex-A application processor, NVIDIA GPU, deep-learning accelerator (DLA) on newer generations, NVDEC and NVENC video blocks, ISP, and LPDDR memory. The customer designs a carrier board that breaks out the SoM connector. NVIDIA supplies a reference carrier (the developer kit); production uses a custom carrier from a third-party integrator or in-house design.

NVIDIA Jetson at a glance
Vendor
NVIDIA Corporation
Founded
1993, Santa Clara, California
Product line
Jetson — edge AI System-on-Module family, 2014 to present
Current generations
Orin (2022 to present); Thor (2024 to 2025 ramp)
Flagship part
Jetson AGX Thor — 14-core Neoverse-derived A720AE + Blackwell GPU, 2000 TOPS INT8
Entry part
Jetson Orin Nano 4 GB — 6-core Cortex-A78AE + Ampere GPU, 20 TOPS INT8
Software stack
JetPack SDK = Linux for Tegra + CUDA + cuDNN + TensorRT + DeepStream + Isaac
External links
developer.nvidia.com/embedded, NGC catalog, Jetson Linux releases

Family timeline

The Jetson family started with the Tegra K1 module in 2014. Each generation tracks NVIDIA's discrete-GPU architecture by roughly two years. The CPU side has migrated from off-the-shelf Cortex-A to NVIDIA's custom Carmel core and now to Cortex-A78AE and the Neoverse-derived A720AE — automotive-grade cores with safety-island features.

GenerationYearCPUGPUINT8 TOPSNotable use cases
TK120144x Cortex-A15Kepler 192 SPn/aDrone vision research, early robotics
TX120154x Cortex-A57Maxwell 256 SP1Skydio drones, NVIDIA Shield TV
TX220172x Denver + 4x Cortex-A57Pascal 256 SP1.3Industrial robotics, smart cameras
Xavier NX20196x Carmel (A75-class custom)Volta 384 SP21Mid-tier robots, smart retail
AGX Xavier20188x CarmelVolta 512 SP32Autonomous mobile robots, medical imaging
Orin Nano20226x Cortex-A78AEAmpere 1024 SP20 to 40Cost-sensitive edge AI, ROS robots
Orin NX20228x Cortex-A78AEAmpere 1024 SP70 to 100Vision pipelines, drones, AGV
AGX Orin202212x Cortex-A78AEAmpere 2048 SP200 to 275Full-stack autonomy, multi-camera, large LLMs
AGX Thor2024 to 202514x A720AE (Neoverse-derived)Blackwell2000Humanoid robotics, autonomous driving

The "AE" suffix marks the automotive-enhanced variant — dual-core lock-step, ECC on L1 and L2 caches, fault-detection / safety-island architecture suitable for ISO 26262 ASIL B (and, with appropriate system design, ASIL D). Jetson is the platform of choice in AV development not because the GPU is uniquely fast but because the CPU complex carries the safety-certification path that consumer-grade ARM SoCs do not.

What you actually get on a Jetson

The Jetson value proposition is not just the GPU. Every current Jetson SoM has:

  • CPU cluster — 6 to 14 cores of Cortex-A running Ubuntu 22.04 LTS via NVIDIA's Linux for Tegra (L4T) distribution with NVIDIA kernel modules baked in.
  • GPU — the headline accelerator. CUDA-compatible from day one; same nvcc, cuDNN, cuBLAS, and Triton inference server as a discrete NVIDIA GPU.
  • DLA — fixed-function INT8 inference engine on Xavier and Orin. Programmed via TensorRT with --useDLACore. Two DLAs on AGX Xavier and AGX Orin; zero on Orin Nano.
  • NVDEC / NVENC — hardware video decode and encode, H.264, H.265, AV1 (Orin onwards). Multi-stream 4K30 decode is the typical motivation.
  • ISP — raw Bayer MIPI-CSI camera input with demosaic, white balance, exposure, denoise, gamma. Camera-vendor partnerships cover Sony, OmniVision, onsemi.
  • Tensor cores — matrix-multiply-accumulate units inside the GPU since Volta. Used by TensorRT for INT8, FP16, FP8, INT4. Drive the published TOPS numbers.
BlockOrin Nano 8 GBOrin NX 16 GBAGX Orin 64 GBAGX Thor
CPU cores681214
CPU clock1.5 GHz2.0 GHz2.2 GHz2.6 GHz
GPU SP cores1024 (Ampere)1024 (Ampere)2048 (Ampere)Blackwell
Tensor cores323264Transformer
DLA count022Integrated
Memory8 GB LPDDR516 GB LPDDR564 GB LPDDR5128 GB LPDDR5X
Peak INT8 TOPS401002752000
Power envelope7 to 15 W10 to 25 W15 to 60 Wup to 130 W

JetPack — the software stack

JetPack is NVIDIA's bundled software distribution. Each release pairs with a specific L4T version and supported hardware:

  • L4T — Ubuntu 22.04 LTS plus NVIDIA kernel modules, GPU driver, multimedia codecs, BSP. U-Boot on older generations, Cboot or UEFI on Orin and Thor.
  • CUDA Toolkit — same CUDA as workstations. JetPack 6 ships CUDA 12.x.
  • cuDNN — NVIDIA's deep-learning primitive library.
  • TensorRT — the inference engine, taking ONNX, torchscript, or SavedModel and producing an optimised engine with INT8 calibration, layer fusion, and kernel auto-tuning.
  • DeepStream — GStreamer-based video-analytics pipeline framework for multi-stream RTSP capture, decode, batched inference, tracking, output.
  • Isaac — robotics platform bridging to ROS 2 with Isaac Sim and robotics primitives.
  • Container Runtime — nvidia-docker and NGC catalog.

The cadence is roughly twice a year for Orin, with longer support on LTS releases.

Why Jetson dominates serious edge AI

Plausible alternatives exist — Qualcomm RB5/RB6, AMD Versal AI Edge, Intel Movidius / Habana, Hailo-8 paired with a Raspberry Pi 5 or i.MX 8M, Coral Edge TPU. Jetson wins consistently in production for three reasons that have little to do with peak TOPS.

First, software lock-in favours NVIDIA. Models are trained on NVIDIA GPUs because that is what is available in cloud, research, and workstations. Moving to a Hailo or Qualcomm or Coral requires conversion, re-quantisation, often re-training, and operator-coverage limits. Moving to Jetson means running TensorRT on the saved checkpoint and accepting a few percent INT8 accuracy loss. The first takes weeks; the second takes hours.

Second, platform breadth. A robotics pipeline is RTSP ingest from N cameras, NVDEC decode, ISP, inference, CUDA post-processing for SLAM or tracking, NVENC encode for storage, network output. On Jetson all of this runs on one chip with NVIDIA-supported plumbing. On alternatives most pieces are separate vendor responsibilities.

Third, tooling maturity. CUDA is a stable, documented API a decade old. TensorRT is nearly as mature. Documentation is good; known bugs have findable workarounds. Alternatives are newer and shifting more.

Pricing tiers and where each fits

ModuleApprox. unit priceTarget deployments
Orin Nano 4 GB$199Hobbyist, education, low-cost industrial vision
Orin Nano 8 GB$299Cost-sensitive vision pipelines, smart cameras
Orin NX 8 GB$399Mid-tier robotics, drones, advanced smart cameras
Orin NX 16 GB$599Multi-camera ingest, ROS-based mobile robots
AGX Orin 32 GB$999Mobile robotics, medical imaging carts
AGX Orin 64 GB$1,999Autonomous vehicles development, surgical robots
AGX Thor 64 GB$3,000+ (early)Humanoid robotics, automotive Level 4 development

Volume pricing is materially lower than the developer-kit list prices above. Customers on multi-thousand-unit programmes typically see 30 to 50 percent discounts.

Cross-references