AI / NPU-Focused SBCs

The "I want a tiny Linux box that runs vision models locally" category. The right pick depends on framework: PyTorch / TensorRT → Jetson; TensorFlow Lite → Coral; ONNX → Hailo; turnkey camera AI → Pi 5 + AI HAT+ or AI Camera. For Pi-side accessory ML hardware see HATs & Accessories; for non-AI SBCs see Alternative SBCs; for cloud / self-hosted LLMs see Self-Hosted AI.

NVIDIA Jetson

The de-facto edge AI SBC. CUDA, TensorRT, the full PyTorch stack — same APIs as the data center.

• ★ Jetson Orin Nano Super (8GB) — refreshed 2024 SKU; ~67 TOPS Sparse INT8, $250. The current sweet spot for hobbyist edge AI: real CUDA + TensorRT, runs YOLOv8 / SAM / small LLMs at usable rates.
• ★ Jetson Orin NX (8/16GB) — module + carrier; ~100-157 TOPS; the "I'm building a product" tier. Need a third-party carrier (Seeed reComputer, Yahboom, Antmicro).
• Jetson AGX Orin (32/64GB) — up to 275 TOPS, real M.2 SSD slot, 8K decode. Workstation-class edge AI.
• Jetson Nano (original) — 472 GFLOPS, 4 GB. Officially EOL'd 2023, software stack frozen at JetPack 4.6 / Ubuntu 18.04. Cheap on eBay but a software dead end. Only buy if you specifically need the old stack.
• Jetson Xavier NX — older 21-TOPS module; used market only.

License: NVIDIA's L4T BSP is permissive-ish; CUDA / TensorRT are closed but free for use. Hardware is mid-priced; the software is the moat.

Google Coral (TPU)

Edge TPU runs TensorFlow Lite integer-quantized models at 4 TOPS @ 2W. Limited model zoo but unbeatable performance-per-watt for the right workload.

• ★ Coral USB Accelerator — plugs into a Pi 4/5 over USB 3; the simplest path. Used heavily in Frigate setups.
• Coral M.2 Accelerator A+E key — plugs into a Pi 5 NVMe HAT or Mini-PC.
• Coral Dev Board / Mini — full SBC (NXP i.MX 8M + TPU). Mostly superseded by "Coral USB on a Pi 5" combinations.
• Coral PCIe Accelerator (Dual / Single) — for x86 servers / NVRs.

Reality: Coral is excellent for Frigate object detection and similar fixed-graph TFLite workloads. For anything that needs PyTorch or a recent transformer model, look at Jetson or Hailo.

Hailo

Israeli AI accelerator; 26 TOPS @ 2.5W; ONNX-based toolchain; growing presence in edge boxes.

• ★ Hailo-8L (13 TOPS) — the chip inside the Pi AI HAT+ / AI Kit, see HATs & Accessories.
• Hailo-8 (26 TOPS) — full chip; available as M.2 (M.2 2280 B+M key) and mPCIe modules.
• Hailo-10H — generative-AI-focused successor (2024 announce); samples in 2026.

Toolchain: Hailo Dataflow Compiler converts ONNX → HEF; HailoRT is the runtime. Open-source-ish (binaries free for use, source mostly closed). Better PyTorch story than Coral; rougher than Jetson.

Raspberry Pi-native AI hardware

Reasonably new; getting the most out of "I have a Pi 5" without buying a Jetson.

• ★ Pi AI Kit / AI HAT+ (Hailo-8L 13 TOPS) — official Pi 5 HAT, ~$70-110, plug-and-play with rpicam-apps and picamera2. The cleanest "add ML to your Pi" path in 2026.
• ★ Pi AI Camera (Sony IMX500) — camera module with on-sensor AI inference (3.1 TOPS); object detection happens before pixels hit the host. ~$70. Lower power than the AI HAT+, model zoo more limited. See Camera Modules.
• Coral USB on a Pi — older but still totally valid for Frigate.

Other NPU-equipped SBCs

NPUs on Rockchip / Amlogic / SpacemiT chips are real but the toolchains are weaker than Jetson / Hailo.

• Khadas Edge2 / VIM4 — RK3588S / A311D2 with vendor NPU; rknn-toolkit converts ONNX → vendor format. Works but rknn is a quirky stack.
• Radxa ROCK 5B / Orange Pi 5 — same RK3588 NPU at 6 TOPS; same rknn-toolkit story. There's an active community wrapping rknn-llm, getting Llama 3 8B running at ~5 tok/s.
• BeagleY-AI — TI AM67A with C7x DSP for ML; TIDL toolchain. Niche.
• Banana Pi BPI-F3 (SpacemiT K1) — RISC-V with 2 TOPS BF16; very early ecosystem.

Frigate-specific recommendations

Frigate (NVR with object detection) is the killer app for edge AI in homelabs. Setup-wise:

• ★ Pi 5 + Coral USB or AI HAT+ — easy, low-power, runs detection on a few cameras.
• ★ Mini-PC with Coral USB — the Frigate docs' top recommendation; a Beelink + Coral handles 8+ cameras at full FPS for ~$300 total.
• Jetson Orin Nano Super — overkill for pure detection, great if you want to add SAM / classification / LLM-narrated alerts.

Pricing rough guide (street, May 2026)

• Coral USB: ~$60
• Pi AI Kit (Hailo-8L): ~$70
• Pi AI Camera: ~$70
• Jetson Orin Nano Super dev kit: ~$250
• Hailo-8 M.2 module: ~$170
• Jetson Orin NX 16 GB module: ~$700 (+ carrier)
• Jetson AGX Orin 64 GB dev kit: ~$2,000

Pick this if…

• First-time edge AI on a Pi: Pi AI Kit (Hailo-8L) or Coral USB.
• Frigate with 4-8 cameras: Coral USB on a mini-PC or Pi 5.
• Real PyTorch / TensorRT on a board: Jetson Orin Nano Super.
• Building a product: Jetson Orin NX module + custom carrier, or Hailo-8 M.2 + your SBC of choice.
• Camera with on-sensor inference: Pi AI Camera (IMX500).
• Already on RK3588: rknn-toolkit works; expect to invest in tooling.
• You're using cloud APIs and it's fine: keep using cloud APIs. Edge AI is for cost, latency, privacy, or air-gap — if none apply, skip.