Boards

Pick a board to see every measured model run, current prices and sources.

Jetson AGX Orin

The big Orin dev kit. Not cheap, but a useful ceiling for edge numbers.

2 runs2 modelsbest 163.9 tok/s

Jetson AGX Thor

NVIDIA's Blackwell-based robotics dev kit, the top of the edge range.

4 runs4 modelsbest 133 tok/s

Jetson Orin NX 8GB (Super mode)

Orin NX 8GB module with JetPack 6.2 Super mode.

7 runs7 modelsbest 69.8 tok/s

Jetson Orin NX 16GB (Super mode)

Orin NX 16GB module with JetPack 6.2 Super mode.

7 runs7 modelsbest 68.8 tok/s

Jetson Orin Nano Super

NVIDIA's small CUDA board in its JetPack 6.2 "Super" (25W / MAXN SUPER) power mode.

21 runs12 modelsbest 64.5 tok/s

RK3588 boards (NPU)

The 6-TOPS NPU in Orange Pi 5, Radxa ROCK 5, Indiedroid Nova, ODROID-M2 and friends, driven by RKLLM.

25 runs24 modelsbest 62.16 tok/s

Jetson Orin NX 16GB (original mode)

Orin NX 16GB module in its original power mode.

7 runs7 modelsbest 59.5 tok/s

Jetson Orin Nano (original / mode not stated)

Orin Nano results in the original power mode, or where the source did not state the mode.

16 runs13 modelsbest 54.8 tok/s

Jetson Orin NX 8GB (original mode)

Orin NX 8GB module in its original power mode.

7 runs7 modelsbest 51.5 tok/s

RK3576 boards (NPU)

Rockchip's mid-range SoC with an NPU, found on boards like the Radxa ROCK 4D.

58 runs20 modelsbest 51.22 tok/s

M5Stack AI Pyramid (AX8850)

M5Stack's AX8850-based AI box.

6 runs5 modelsbest 37.11 tok/s

M5Stack LLM8850 M.2 (AX8850)

An AX8850 accelerator on an M.2 card, for adding an NPU to another host.

7 runs7 modelsbest 27.05 tok/s

Pi 5 cluster (distributed-llama)

Several Pi 5 boards splitting one model over Ethernet with distributed-llama.

2 runs1 modelsbest 13.68 tok/s

M5Stack Module LLM (AX630C)

A tiny 54mm module with an Axera AX630C NPU, running StackFlow.

11 runs6 modelsbest 12.52 tok/s

Raspberry Pi 5 (CPU)

The default maker board. LLMs run on the four Cortex-A76 CPU cores with llama.cpp or Ollama.

34 runs17 modelsbest 11.73 tok/s

Pi 5 + AI HAT+ 2 (Hailo-10H)

A Hailo-10H NPU with its own 8GB RAM on a HAT, running Hailo-compiled models through hailo-ollama.

18 runs6 modelsbest 8.7 tok/s

Arduino UNO Q

Arduino's Linux + microcontroller board. Results so far are CPU-only.

1 runs1 modelsbest 5.5 tok/s

Get the monthly board report

New boards, new runtimes, new numbers. One email a month, no hype.