NVIDIA Jetson AGX Thor vs Orin: Full Series Comparison

The NVIDIA Jetson ecosystem has expanded rapidly, with Orin modules powering everything from entry-level AI applications to industrial robotics. Now, the arrival of Jetson Thor brings a massive leap in compute and efficiency. In this article, we’ll break down the Jetson Orin family and compare it with the new Thor series. To make things clear, we’ll start with a detailed comparison table and then walk through each module series.

The architectural shift from Jetson AGX Orin to Jetson AGX Thor is not a simple generational upgrade but a transition to a fundamentally new compute stack. Orin is built on NVIDIA’s Ampere GPU architecture paired with Arm Cortex-A78AE CPUs, while Jetson AGX Thor introduces the newer Blackwell GPU architecture alongside Arm Neoverse-V3AE cores. This change enables Thor to deliver significantly higher AI throughput and efficiency, with NVIDIA reporting more than 7.5× higher AI compute and around 3.5× better energy efficiency compared to Orin, positioning it for next-generation robotics and physical AI workloads. Here are the key differences:

• Performance: Jetson Thor uses the Blackwell architecture, which offers up to 2070 TFLOPS (FP4 sparse). Compared to Jetson AGX Orin's peak of ~275 TOPS (INT8), Jetson Thor offers more compute power.
  
• Use Cases: The ideal use cases differ as Jetson Thor brings more power to the table meanwhile Jetson AGX Orin is ideal for standard mobile robots and entry-level automation.
  
• Memory & I/O: Jetson AGX Thor supports up to 128GB LPDDR5X, which is double Orin's 64GB max, and PCIe Gen 5.
  

The GPU is the primary driver of the performance gap between the two platforms. Jetson AGX Orin integrates an Ampere GPU with 2048 CUDA cores and 64 Tensor cores, delivering up to 275 TOPS of AI performance depending on configuration.

In contrast, Jetson AGX Thor moves to a Blackwell GPU with 2560 CUDA cores and 96 next-generation Tensor cores, alongside a dramatic increase in AI compute that reaches into the thousands of TOPS or equivalent FP4 throughput levels.

This increase is not just about raw numbers; newer Tensor cores and architectural improvements enable more efficient execution of modern AI workloads such as transformers, multimodal models and large-scale perception pipelines, which are increasingly common in robotics and edge AI systems.

The CPU evolution is maybe what plays the most critical role in real-world system performance. Jetson AGX Orin’s 12-core Cortex-A78AE CPU is designed for deterministic, safety-capable workloads typical in robotics, including sensor handling and control loops.

Jetson Thor upgrades this to a 14-core Arm Neoverse-V3AE CPU, delivering higher clock speeds, larger cache structures, and improved multi-threading capabilities.

In practical deployments, this translates to better handling of concurrent tasks such as sensor fusion, path planning, and middleware execution (e.g., ROS 2), reducing CPU bottlenecks and enabling more consistent utilization of the GPU in complex, multi-modal AI pipelines.

Memory is increasingly a limiting factor for edge AI, especially with the rise of large models and multi-sensor systems. Jetson AGX Orin supports up to 64GB of LPDDR5 memory with bandwidth around 200 GB/s, which is sufficient for most perception and vision workloads.

Jetson Thor significantly expands this with up to 128GB of LPDDR5X memory and bandwidth exceeding 270 GB/s, enabling faster data movement and larger on-device models.

This improvement is critical for applications such as:

• Real-time multi-camera processing
• High-resolution sensor fusion
• Running large language models
• Deploying vision-language models locally

 As these are all operations where both memory capacity and bandwidth directly impact latency and throughput.

Power and thermal characteristics are one of the clearest practical differences between Jetson AGX Orin and Jetson AGX Thor.

Jetson AGX Orin is designed for efficiency, with a configurable power envelope typically ranging from 15W to 60W, making it suitable for embedded and battery-powered systems where thermal constraints are strict.

In contrast, Jetson AGX Thor operates in a much higher power range, roughly 40W to 130W depending on configuration, reflecting its significantly higher compute capability. Despite this, Thor delivers around 3.5× better performance per watt than Orin, meaning that while absolute power consumption increases, efficiency relative to compute improves substantially. In practice, this creates a clear trade-off: Jetson AGX Orin is easier to integrate into compact, thermally constrained designs, while Jetson AGX Thor requires more robust cooling and power delivery but enables far more demanding AI workloads.

Choosing between Jetson AGX Thor and Jetson AGX Orin ultimately comes down to aligning compute requirements with system constraints such as power, thermal design, and deployment maturity. While Thor represents a major leap in raw performance and is aimed at next-generation AI workloads, Orin remains a highly capable and widely adopted platform with a strong balance between performance, efficiency, and ecosystem maturity. The decision is less about which is “better” overall and more about which fits the specific demands of your application, especially in terms of model complexity, real-time requirements, and hardware limitations. 

Orin NX and Orin Nano continue to play an important role in the Jetson ecosystem by addressing lower-power and cost-sensitive segments. These modules offer multiple power modes, typically ranging from around 10W up to 40W depending on configuration, allowing developers to optimize performance within tight energy budgets. They remain highly relevant for compact systems such as drones, small robots, and edge devices where size, weight, and thermal constraints are more restrictive, and where the full capabilities of AGX-class modules are unnecessary.

The Jetson Orin family gives developers a wide range of compute options, scaling from low-power Nano modules to industrial-grade AGX Orin. With Jetson Thor, NVIDIA is redefining the upper limits of AI performance for robotics and edge computing. The choice now depends on your application’s power budget, size constraints, and compute needs.