Almost every "best SBC for AI" article you will find ranks boards by CPU speed and then mentions the NPU in one line. That is backwards. For inference work the NPU is the whole point, and the three things that decide whether a board is usable are TOPS at INT8, whether a real toolchain exists, and how much RAM you can put behind the model. Here is how the boards in our catalog actually compare on those three.
Quick Picks
- Best NPU board overall: Radxa ROCK 5B — 6 TOPS, RKNN toolkit, up to 16GB, PCIe 3.0 x4 for model storage, from $59.
- Best for camera and NVR work: FriendlyElec NanoPi R6S — 6 TOPS, dual 2.5GbE, headless, $85 to $105.
- Best desktop-class NPU board: FriendlyElec NanoPC-T6 — 6 TOPS with LPDDR5 and dual 2.5GbE, $109 to $139.
- Most RAM behind the NPU: Orange Pi 5 Plus — 6 TOPS with up to 32GB LPDDR5.
- Only RISC-V option: Sipeed LicheePi 4A — 4 TOPS, TH1520, early toolchain, $75 to $130.
What a TOPS Number Actually Tells You
TOPS is trillions of operations per second, and it is only comparable when everyone quotes the same precision. Our catalog stores dense INT8 figures, which is the number that matches what you get running a quantized model. Vendors that quote INT4 or sparse numbers are not lying, they are just not answering the same question, and a board advertised at 12 TOPS sparse may be the same 6 TOPS dense part as its neighbour.
What TOPS does not tell you is whether the accelerator is reachable. A vendor NPU with no exporter, no runtime, and no quantization tooling is a spec sheet entry, not a feature. That is the real axis these boards separate on.
The Toolchain Question
Rockchip's RK3588 and RK3588S boards all expose a 6 TOPS NPU through RKNN, which is a genuine working stack: an ONNX and PyTorch converter, an INT8 quantizer, and a runtime with Python and C APIs. Object detection, pose estimation, classification and segmentation all run at usable frame rates. You will find community model zoos and working YOLO exports.
The Sipeed LicheePi 4A is the contrast. Its T-Head TH1520 has a 4 TOPS NPU and up to 16GB of LPDDR4X, which is more memory than almost any other RISC-V board, and on paper it is the closest RISC-V has come to an RK3588-class part. In practice the toolchain is early. ONNX and TVM paths exist and are being ported, but they are nothing like RKNN in maturity. We rate it 6.9 and the reason is entirely software, not silicon.
Buy the LicheePi 4A because you want to be early on RISC-V AI. Do not buy it because you need an inference box that works next week.
Matching the Board to the Workload
Camera and NVR pipelines
The NanoPi R6S is the specialist here. Same RK3588S silicon and 6 TOPS NPU, but laid out as a router board: dual 2.5GbE, no HDMI, no GPIO, no MIPI camera connectors, eMMC and microSD only, 90 x 64mm. That sounds like a list of omissions until you are running Frigate with several IP cameras, at which point the two 2.5GbE ports and the headless form factor are exactly right and the missing HDMI is irrelevant. It also routes, so one box can be firewall and NVR at once. $85 for 4GB, $105 for 8GB.
General vision and development
The Radxa ROCK 5B is the board to develop on. 6 TOPS, two MIPI-CSI connectors and one MIPI-DSI, PCIe 3.0 x4 to an M.2 slot so models and clips live on NVMe rather than an SD card, 2.5GbE, and dual HDMI. At $79 for 8GB it is the cheapest way to have every piece of the pipeline on one board.
Larger models
Where boards run out is memory, not TOPS. The Orange Pi 5 Plus goes to 32GB LPDDR5, which matters the moment you are holding a model, a framebuffer and a container runtime at once. The NanoPC-T6 caps at 16GB but pairs it with LPDDR5 and dual 2.5GbE, which makes it the better pick when inference is one job on a box that is also serving something.
Comparison Table
| Board | NPU | SoC | Max RAM | Toolchain | Camera inputs | Price |
|---|---|---|---|---|---|---|
| Radxa ROCK 5B | 6 TOPS | RK3588 | 16GB LPDDR4X | RKNN | 2x MIPI-CSI | $59 to $99 |
| FriendlyElec NanoPi R6S | 6 TOPS | RK3588S | 8GB LPDDR4X | RKNN | None | $85 to $105 |
| FriendlyElec NanoPC-T6 | 6 TOPS | RK3588 | 16GB LPDDR5 | RKNN | 4-lane MIPI-CSI | $109 to $139 |
| Orange Pi 5 Plus | 6 TOPS | RK3588S | 32GB LPDDR5 | RKNN | MIPI-CSI | Varies |
| Sipeed LicheePi 4A | 4 TOPS | TH1520 RISC-V | 16GB LPDDR4X | Early ONNX/TVM | 2x MIPI-CSI | $75 to $130 |
| Raspberry Pi 5 (8GB) | None | BCM2712 | 16GB LPDDR4X | CPU only | 2x 4-lane MIPI | $80 |
The Pi 5 is in the table for one reason: it has no NPU at all. If inference is your workload, that is the line that decides it.
Winner: Radxa ROCK 5B
The Radxa ROCK 5B wins because it is the cheapest board that gives you all three things at once: 6 TOPS through a toolchain that works, 16GB of RAM available, and PCIe 3.0 x4 so the storage under the model is not a microSD card. At $79 for 8GB it is less than a Pi 5 that cannot do the job at all.
Take the NanoPi R6S instead if the workload is cameras on a network and you will never plug in a monitor. Take the Orange Pi 5 Plus if 16GB is your ceiling problem. Take the LicheePi 4A only if RISC-V itself is the reason you are buying.
You can filter the whole single-board computer catalog by NPU presence if you want to check this yourself. For the wider question of what to buy when the Pi 5 is not enough, see Raspberry Pi 5 Alternatives, and if storage bandwidth is your constraint rather than inference, see Single-Board Computers with a PCIe Slot.
Related on Consumercarts
From the Consumercarts catalog: