Jetson Arena
The arena for model setups on Jetson devices — from a single model to a whole pipeline.
First targetJetson Thorrealtime voice loop: VAD → STT → LLM → TTS
Jetson Arena answers one question: how does a given Jetson device handle a given setup? A single model, a mixture of models, or a whole pipeline — any configuration, measured as it actually runs, on the device everything has to share.
Every published run will carry three things: its latency, end to end and per stage; its memory signature on unified memory; and an honest read of its quality — alongside the Docker recipe that reproduces it.
Nothing is measured yet. The first target is chosen, and the method is public before any number is — so that when results land, they can be held to it.
Jetson Thor
A realtime voice loop, running whole on one board: speak, and hear the reply.
- VADvoice activity detection
- STTspeech to text
- LLMlanguage model
- TTStext to speech
Four stages contending for one pool of unified memory, in a loop a person can feel. It is exactly the kind of setup that only makes sense measured whole — and none of it is measured yet. Next in line: Jetson AGX Orin on JetPack 7.2, then Jetson Orin Nano Super 8GB.
Three axes, stated before any number exists
Latency
End to end, and per stage
The wall-clock story of one turn: from the end of speech to the first audible reply, and where inside the pipeline that time goes. Every published run will carry the full breakdown, not a single headline figure.
Memory signature
On unified memory
CPU and GPU share one pool on a Jetson, so the stack's whole footprint is what matters. Runs will record what each stage holds resident, and what remains for everything else on the board.
Quality
Honest, per stage
A loop that replies fast but mishears the question has not won anything. Each stage will be scored against its own task, and every trade-off made to fit the device will be stated with the run.
Every run will ship its recipe
Every published run will come with the Docker recipe that reproduces it — the exact containers, models, and flags, building on the jetson-containers family. Own the device, rerun the arena: a result you cannot reproduce is a result you should not trust.