Skip to content

Product1 publisher3 min readPublished Updated

Nvidia's entry-level robot brain doubles to 78 TOPS and keeps its 8 GB ceiling

Jetson Orin Nano 2 takes the entry seat in edge robotics to 78 trillion operations per second, but not until 2027, and the memory figure is what decides which small models ship.

The Product Desk · Product desk

What happened

  • Nvidia announced the Jetson Orin Nano 2, an entry-level robotics module aimed at running frontier-class models on the device itself.
  • The board is specified at 78 trillion operations per second, 8 gigabytes of memory and an eight-core CPU.
  • Cognex, Doosan Bobcat, Matic Robotics and Wing Aviation are already designing it into products.

Compiled by The Product DeskSomething wrong?How this is made

Why it matters

  • capability Perception that had to survive a radio hop to a datacentre can now be budgeted as local latency, which is the only version of collision avoidance that a certifier will look at.
  • constraint Model choice on this board is still governed by the memory pool, so teams get faster inference and room for a second network rather than permission to load a larger one.
  • cost Because the efficiency figure is quoted for a power mode rather than a matched workload, the buyer carries the measurement risk when converting it into battery minutes or thermal margin.
  • decision Anyone freezing a robotics bill of materials in the next two quarters has to choose between shipping on silicon that is about to be halved in relative terms and waiting on a part with no firm date...

The figure that decides which model actually loads is 8 gigabytes [3], and that is the one the announcement leaves alone. Compute doubles over the previous generation [2] to 78 trillion operations per second [3]; memory is listed at 8 GB with no stated increase [15]. Divide one by the other and the part carries about 9.8 TOPS for every gigabyte it can hold [16]. On a drone, the weights, the attention cache, the camera buffers and the operating system all draw on that same pool, so a doubling of arithmetic throughput buys frame rate and headroom for a second network, not a bigger model than the old part could fit.

The power claim needs the same care. Nvidia says the module draws 40% less power than its predecessor in 15-watt mode [4]. Set that against the doubled compute and you get roughly 3.3 times the work per watt [13], but only if the 40% is measured at matched throughput, which the source does not say [4]. For Wing Aviation, whose head of perception Dinuka Abeywardena says drone delivery depends on AI that can give fast, reliable understanding of the real world [8], that ambiguity is denominated in flight minutes.

The latency argument is the strongest part of the pitch, and it is not about model quality. Nvidia's case is that vision runs on the airframe instead of going out over wireless to a datacentre and back [10], and Wing is looking at the module for autonomous navigation and avoidance in its delivery fleet [7]. Avoidance is the workload where a radio round trip is not a cost to be optimised but a reason the function cannot exist.

Then there is the calendar. The module and developer kit reach general availability in the first half of 2027 [5], four to ten months after the August 2026 announcement [14][11]. Deepu Talla, Nvidia's vice president of robotics and edge AI, argues that small and medium frontier models have already caught last year's largest ones on accuracy [9], which is the premise the hardware is specced against. If he is right, the models keep compressing across that gap and the 78 TOPS looks better on arrival than it does on paper now. The 8 GB does not move with them.

Four named partners are already designing it in: Cognex, Doosan Bobcat, Matic Robotics and Wing [6]. What they are committing to is a socket where the entry-level floor rises from roughly 39 TOPS to 78 [12][2]. Anyone specifying a 2027 robot on the outgoing part is now buying half the compute for the same slot and the same power budget, which is a harder line to defend to a customer than it was last week.

What to watch

  • Whether Nvidia publishes matched-throughput power numbers, or a memory option above 8 GB, before the developer kit ships.
  • Whether the first-half 2027 availability window holds, and which of the four named partners moves from evaluation to a shipping unit.
  • Whether Wing commits the module to a certified delivery airframe rather than a perception testbed.
Loading claim ledger
Loading source directory links
Loading share composer
Loading topic controls
Loading related stories