Jetson Orin Nano 2 Hits 78 TOPS on 40% Less Power

Twice the Inference in the Same Footprint

Nvidia has detailed the Jetson Orin Nano 2, the next entry-level module in the Jetson line, with 78 TOPS of AI compute against the 67 sparse TOPS of the Orin Nano Super it replaces. LinuxGizmos reported the specifications on August 26, 2026. The module and its developer kit are expected in the first half of 2027.

• 78 TOPS of AI performance, roughly double the inference throughput of the Orin Nano Super in the same physical footprint
• 8GB of LPDDR5X with up to 120 GB/s of memory bandwidth
• Eight Arm Cortex-A78 CPU cores handling the non-accelerated half of the workload
• In a 15W mode it matches its predecessor's performance using about 40% less power, which is the more interesting of the two numbers

Why the 40% Power Figure Beats the TOPS Figure

Peak TOPS is the number that goes in the headline, and it is the number that matters least to most people building with these modules. Entry-level Jetson boards end up in drones, inspection robots and battery-powered vision systems, and in all three the binding constraint is watts, not ceiling performance.

Framed as an efficiency result, the Orin Nano 2 says something more useful: take the workload you are already running on an Orin Nano Super, keep the same throughput, and give back about 40% of the power. On a delivery or inspection drone, that converts directly into flight time. On a battery-powered camera node, it converts into either a longer service interval or a smaller battery and a lighter enclosure.

The 120 GB/s of LPDDR5X bandwidth is the supporting detail that makes the doubled throughput plausible. Vision pipelines on small modules are frequently memory-bound rather than compute-bound — you can have all the TOPS you like and still stall waiting on frame buffers. Pairing the accelerator uplift with the bandwidth to feed it is what turns a specification into a real-world doubling.

Where Does This Sit in the Jetson Range?

At the bottom, which is exactly the point. The upper end of the family has moved a long way in the past year, from the Jetson T3000 bringing Blackwell edge AI to small robots to carrier designs like the reComputer Mini J501 packing 275 TOPS into a 119mm cube. Those are for machines with a real power budget.

The Orin Nano tier is where hobbyist robotics, university labs and small commercial products actually live, largely because the Orin Nano Super landed at $249 in late 2024 and made a capable Jetson affordable for the first time. Nvidia has not announced pricing for the Orin Nano 2, which is the single most important unknown here — the module's usefulness to this audience is almost entirely a function of what it costs.

What 8GB Means for On-Device Models

Eight gigabytes is the same memory the Orin Nano Super carried, and it is a real ceiling. It comfortably runs vision models, multi-camera detection and tracking pipelines, and small language models in quantized form. It will not host anything approaching the open-weight models people run on desktop hardware, and it is not meant to.

For the robotics and drone work this module targets, that trade is usually the right one. These systems need fast, deterministic perception on a tight power envelope far more than they need a large model, and our broader edge AI hardware coverage keeps landing on the same conclusion: the winning specification at this tier is performance per watt, and the Orin Nano 2 is aimed squarely at it.

The first half of 2027 is a long wait, and Nvidia's announced-to-shipping gaps have run long before. But the shape of the product is clear, and it is the shape the category needs.

Sources: LinuxGizmos — August 26, 2026; StorageReview — August 2026.