Local inference for robotics, smart home and content creation, where latency, privacy or connectivity rules out sending data to a server.
Cloud inference adds latency, a per-request cost, and a dependency on a network that may not be there. For robotics control loops and for anything handling video of people, the round trip is often the problem rather than the model.
| Local first | The K3 is specified at 60 TOPS INT4 from eight dedicated A100 AI cores, with LPDDR5 up to 32 GB at 51 GB/s. That is enough memory to hold quantized models that would otherwise be split across a network. |
|---|---|
| Mainline software | RVA23 compliance means the K3 runs Ubuntu 26.04 LTS natively, without out-of-tree patches. Your ML stack is a standard package install, not a vendor fork you have to maintain. |
| Vision and IO | Four MIPI-CSI inputs supporting up to 12 cameras, 4K@180fps decode and 4K@90fps encode. Four TSN-capable Gigabit MACs and eight lanes of PCIe Gen3 for accelerator or NVMe expansion. |
Can it really run a 30B model?
A quantized 30B model, yes — that is what 60 TOPS INT4 and 32 GB of LPDDR5 are for. Token throughput depends on the quantization and the runtime; benchmark your own model before committing to a figure.
Is it a drop-in for Jetson Orin Nano?
The K3 is pin-compatible with Jetson Orin Nano, which makes carrier board migration realistic. Software migration is a separate piece of work.