Resources
Jetson Orin vs. Jetson Thor:
Which Platform Fits Your Workload?
Both run the same JetPack ecosystem. The decision comes down to compute density, power budget, and whether your workload needs generative or agentic ai at the Edge.
By Ceri Nelmes
Tech Marketing Leader & Journalist | Building Brands at the Edge of What’s Next | Connect Tech & CTai Labs
Technical review: Rob Callaghan, Chief Product Officer, Connect Tech Inc.
Key Takeaways
- NVIDIA® Jetson AGX Orin™ delivers up to 275 TOPS at 15 to 60 watts, the right fit for most industrial vision, logistics AMR, and defense payload applications with well-understood compute requirements
- Jetson Thor™ (T5000) delivers 2,070 FP4 TFLOPS at 40 to 130 watts on the NVIDIA Blackwell™ architecture, roughly 7.5x the ai compute of Jetson AGX Orin with 3.5x the energy efficiency
- Jetson Thor is built for humanoid robotics, advanced autonomous mobile robots, and vision-language-action (VLA) model deployment
- Both platforms run on Connect Tech solutions, with BSPs written in-house
- CTai LABS helps teams select, optimize, and validate the right NVIDIA Jetson™ platform for their workload, balancing compute performance, memory, power, thermal requirements, and deployment constraints
- Teams on Rogue (AGX202) or Anvil (ESG620) have a direct migration path to Rogue-T5 (AGX302) or Anvil-T5 (ESG625) for Jetson Thor
Full spec comparison
The table below covers the production specifications for NVIDIA® Jetson AGX Orin™ (64GB) and Jetson Thor™ (T5000), the highest-tier modules in each family (NVIDIA, 2026; ServeTheHome, 2025; HotHardware, 2025).
| Spec | Jetson AGX Orin (64GB) | Jetson Thor (T5000) |
|---|---|---|
| AI Compute | Up to 275 TOPS (INT8 Sparse) | 2,070 FP4 TFLOPS / 1,035 FP8 TFLOPS |
| GPU Architecture | NVIDIA Ampere, 2,048 CUDA cores, 64 Tensor cores | NVIDIA Blackwell, 2,560 CUDA cores, 96 5th-generation Tensor cores |
| CPU | 12-core Arm Cortex-A78AE | 14-core Arm Neoverse-V3AE |
| Memory | 64 GB LPDDR5, 256-bit, 204.8 GB/s | 128 GB LPDDR5X, 256-bit, ~273 GB/s |
| Power Envelope | 15–60 W (configurable) | 40–130 W (configurable) |
| Networking | Up to 10GbE | Up to 4× 25GbE |
| Generative AI / VLA Support | Capable, compute-limited for large multi-model pipelines | Native FP4 Transformer Engine; built for VLA (Isaac GR00T) and multi-model agentic workloads |
| MIG (Multi-Instance GPU) | Not supported | Supported, up to 2 isolated GPU instances |
| Typical Use Case | Industrial vision, logistics AMRs, defense payloads, and compact deployments | Humanoids, advanced AMRs, multi-sensor ISR, agentic and VLA-driven robotics |
| Operating Temperature (Connect Tech carriers) |
−40°C to +85°C | −40°C to +85°C (carrier-dependent) |
| JetPack™ | JetPack 6.x | JetPack 7.x |
Note: Jetson AGX Orin Industrial (used on Rogue-RX/AGX203) delivers up to 248 TOPS at 15–75 W with extended temperature and ECC memory support. Jetson Thor also ships in a lower-tier T4000 variant (up to 1,200 FP4 TFLOPS, 64 GB memory, 12-core CPU) for applications that do not need the full T5000 compute envelope (Connect Tech, 2025; NVIDIA, 2026).
What actually drives the decision
Most platform decisions come down to four questions, and they map cleanly onto the spec table above.
Is the workload generative or agentic?
If the application runs a single perception or detection model, classifying defects, navigating a fixed route, Jetson AGX Orin’s 275 TOPS is generally sufficient. If the application needs to run a vision-language-action model, a world foundation model, or multiple large models simultaneously, generative ai for robot reasoning, multimodal sensor fusion, Jetson AGX Orin becomes the bottleneck. Jetson Thor’s native FP4 Transformer Engine and Multi-Instance GPU (MIG) support exist specifically for this class of workload.
What is the power budget?
Jetson AGX Orin’s 15-to-60-watt range fits battery-powered and thermally constrained platforms where every watt matters. Jetson Thor’s 40-to-130-watt range delivers more headroom but requires a platform that can supply and dissipate that power. For SWaP-limited UAV payloads, Jetson AGX Orin or smaller Jetson Orin™ NX/Nano modules remain the right choice. For ground-based AMRs and humanoid platforms with more available power budget, Jetson Thor’s efficiency gain, 3.5 times over Jetson AGX Orin per the manufacturer’s published figures, can outweigh the higher absolute power draw.
What is the memory requirement?
More memory provides headroom for larger models and more demanding workloads, but it also affects system cost. Jetson AGX Orin configurations support up to 64GB of memory, while Jetson Thor increases that ceiling to 128GB for larger generative ai models, vision-language-action models, and workloads that need multiple models resident in memory. The goal is not to buy the most memory available, but to determine what the workload requires. CTai LABS can optimize models and memory utilization to help teams avoid overprovisioning hardware and paying for capacity they do not need.
What form factor does the deployment require?
Compute requirements must fit within the physical constraints of the deployed system. Jetson Orin spans multiple module sizes and performance levels, from the compact Jetson Orin Nano and Orin NX modules to Jetson AGX Orin, giving teams more options for UAVs, compact robots, embedded vision systems, and other SWaP-constrained deployments. Jetson Thor targets applications that can accommodate a larger, higher-power platform in exchange for substantially more compute and memory. For space- and weight-constrained systems, form factor can rule out a platform before performance becomes the deciding factor.
How many cameras and sensors does the system carry?
Jetson Thor’s networking jumps to up to 4x 25GbE compared to Jetson AGX Orin’s 10GbE, and Connect Tech’s Jetson Thor-ready carriers (Gauntlet, Rogue-T5, Anvil-T5) support GMSL3 in addition to GMSL2 and FPD-Link III. For multi-camera, high-resolution perception systems, Jetson Thor’s I/O headroom is frequently the deciding factor independent of raw compute.
Is this a new build or a migration?
Teams already deployed on Jetson AGX Orin via Rogue or Anvil have a defined upgrade path to Jetson Thor through Rogue-T5 and Anvil-T5, which share enough architectural lineage that the transition is closer to a platform upgrade than a greenfield redesign. See the migration path table below.
Figure 1. Jetson AGX Orin and Jetson Thor platform selection based on workload, power, memory, cost, form factor, I/O requirements, and existing deployment architecture.
Migration path: Rogue and Anvil to Jetson Thor
Connect Tech designed its Jetson Thor-ready carrier lineup, Gauntlet (AGX301), Rogue-T5 (AGX302), and Anvil-T5 (ESG625), as documented direct upgrade paths from the equivalent Jetson AGX Orin hardware. Teams on Rogue or Anvil today are not starting from zero when they move to Jetson Thor.
NVIDIA describes this as a “three-computer” architecture for robotics specifically: training infrastructure, simulation infrastructure, and the on-robot inference computer (NVIDIA, 2025). The same pattern generalizes to most Edge ai systems beyond robotics.
Rogue to Rogue-T5
Rogue (AGX202) and Rogue-RX (AGX203) users moving to Jetson Thor migrate to Rogue-T5 (AGX302), which carries forward the compact form factor and adds GMSL3, FPD-Link III, SDI, and expanded CAN and Ethernet I/O for the higher sensor bandwidth Jetson Thor workloads typically require (Connect Tech, 2026).
Anvil to Anvil-T5
Anvil (ESG620) users move to Anvil-T5 (ESG625), which extends the same vision-intensive feature set, GMSL3, GMSL2, FPD-Link III, and SDI, on a JetPack 7-enabled platform built specifically for multi-camera Jetson Thor deployments (Connect Tech, 2026).
Forge to Gauntlet
There is no direct Jetson AGX Orin predecessor to Gauntlet; Gauntlet was introduced as Connect Tech’s flagship Jetson Thor carrier. For teams sizing up from Jetson AGX Orin, Forge (AGX201) is the closest equivalent in form factor and full-featured I/O, and is the natural reference point when planning a move to Gauntlet for Jetson Thor-class compute (Connect Tech, 2025; Connect Tech, 2026).
Connect Tech publishes a complete Jetson Module Migration Guide covering every supported upgrade path with L4T compatibility tables and I/O change notes. For the step-by-step technical process of porting a deployment to a new Jetson platform, see the x86 to Jetson Migration Guide, which covers many of the same BSP and software porting steps relevant to an Orin-to-Thor migration.
Get help choosing
Platform selection is the highest-leverage decision in an Edge ai project. CTai LABS, as a department of Connect Tech, scopes platform selection against your actual workload, power budget, and sensor requirements as the first step of any engagement, not after the architecture is already locked in.
Book a Demo
Have an Edge ai workload that needs to move from model to deployment? Talk to CTai LABS about your application, performance requirements, data constraints, and deployment environment.
ABOUT THE AUTHOR
Ceri Nelmes
Tech Marketing Leader & Journalist | Building Brands at the Edge of What’s Next |
Connect Tech & CTai Labs
Ceri Nelmes is Head of Marketing at Connect Tech and an experienced technology journalist and digital strategist. She covers the technologies and market shifts shaping embedded computing, Edge AI, Physical ai, robotics, autonomous systems, and the NVIDIA® ecosystem. Working with Connect Tech and CTai LABS subject-matter experts, she turns engineering developments into accurate, useful reporting for developers, technical buyers, business leaders, and the media.
Learn More
Resources and Frequently Asked Questions
Related
Start Reading
See why memory bandwidth, not TOPS, is often a real Edge ai constraint.
Book a Demo
Not sure which module fits your workload?
Talk to a CTai LABS engineer or ai architect.
Sources
Connect Tech. (2025). Rogue carrier board for NVIDIA Jetson AGX Orin (AGX202); Rogue-RX carrier board (AGX203); Anvil rugged system (ESG620).
https://connecttech.comConnect Tech. (2026). Your Jetson Xavier platform is approaching end of life: Migration guide.
https://connecttech.com/jetson-xavier-eol-migration-guide/HotHardware. (2025, August 25). NVIDIA Jetson AGX Thor tested: Blackwell brings physical AI to life.
https://hothardware.com/reviews/nvidia-jetson-agx-thor-developer-kit-hands-onNVIDIA. (2026). Jetson AGX Orin: Next-generation robotics and Edge ai.
https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/NVIDIA. (2026). Jetson Thor: Advanced ai for physical robotics.
https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/NVIDIA. (2026). NVIDIA Blackwell-powered Jetson Thor now available, accelerating the age of general robotics.
https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-roboticsServeTheHome. (2025, August 29). NVIDIA Jetson AGX Thor developer kit hands-on: Blackwell for robotics.
https://www.servethehome.com/nvidia-jetson-agx-thor-developer-kit-blackwell-for-robotics/Frequently Asked Questions
Is Jetson Thor just a faster version of Jetson AGX Orin?
Not exactly. Jetson Thor is built on a different GPU architecture (Blackwell vs. Ampere) with native FP4 quantization support and a Transformer Engine specifically for generative and agentic workloads. It is a higher-compute platform, but the architectural differences mean it is best suited to a different class of application, multi-model, generative, and agentic ai, rather than simply being a faster version of the same workload.
Can I run my existing Jetson AGX Orin software on Jetson Thor without changes?
Most application-level code ports with minimal changes since both platforms run JetPack and CUDA™. TensorRT engines, however, are architecture-specific and must be rebuilt for Jetson Thor’s Blackwell GPU. BSP and driver-level code tied to Jetson Orin-specific hardware features will need carrier-specific updates for the new platform.
Does Jetson Thor support all the same NVIDIA software as Jetson AGX Orin?
Jetson Thor runs JetPack 7 and supports TensorRT™, Isaac™ ROS, DeepStream, and Holoscan, the same core stack as Jetson AGX Orin. Jetson Thor adds support for newer capabilities, including native FP4 inference and Multi-Instance GPU (MIG), that are not available on Jetson AGX Orin’s Ampere architecture.
Why would I choose Jetson AGX Orin over Jetson Thor if Jetson Thor is more powerful?
Power budget and cost. Jetson AGX Orin’s 15-to-60-watt range fits applications, especially battery-powered or thermally constrained ones, where Jetson Thor’s 40-to-130-watt range is not viable. For workloads that do not require generative or multi-model ai, Jetson AGX Orin’s compute is sufficient and the platform is more cost-effective.
What is the difference between Jetson T4000 and T5000?
Both are Jetson Thor modules on the Blackwell architecture. T5000 is the higher-tier variant with up to 2,070 FP4 TFLOPS, 128 GB memory, and a 14-core CPU. T4000 delivers up to 1,200 FP4 TFLOPS with 64 GB memory and a 12-core CPU, for applications that need Jetson Thor-class architecture but not the full T5000 compute envelope.
If I am on Rogue or Anvil today, what is my upgrade path to Jetson Thor?
Rogue (AGX202) and Rogue-RX (AGX203) users move to Rogue-T5 (AGX302). Anvil (ESG620) users move to Anvil-T5 (ESG625). Both transitions carry forward familiar form factors and I/O patterns while adding the camera bandwidth and networking Jetson Thor workloads typically need. Connect Tech’s Jetson Module Migration Guide documents the full transition in detail.
Ready to Build Smarter?
Let’s create the intelligent Edge AI solution that moves your business forward.