FIELD MANUAL · § ALT-CTRL · COMPARE FM 18-DOF / PI5 vs UNO-Q
★ PROJECT HEXAPOD ★ INTERNAL

Pi 5 + AI HAT vs Arduino UNO Q.

TWO WAYS TO BRAIN A HOUSE-CREATURE · LLM-FIRST READ

Both are dual-architecture in spirit, but only the UNO Q has a real MCU on board. The Pi 5 is a pure SBC — vastly more compute, much more RAM, and (with the AI HAT) real NPU horsepower — but you'll need a second chip for hard-real-time servo timing. This sheet is biased toward the question that actually decides it for you: how much LLM can each one run locally?

01 · The two stacks, side by side

OPTION A · MORE BRAIN

Raspberry Pi 5 (16 GB) + AI HAT+

SBC · external MCU sidecar · big LLM ceiling
  • CPU4× Cortex-A76 @ 2.4 GHz · ~2–3× IPC of A53
  • RAM16 GB LPDDR4X (8 GB also available)
  • NPU26 TOPS Hailo-8 via AI HAT+ (CNN-focused; weak on LLM today)
  • StoragemicroSD + NVMe via HAT (PCIe 2.0)
  • Camera2× MIPI-CSI · plus IMX500 "AI Camera" option
  • GPIO40-pin (3.3 V) + PCIe + USB 3 — vast accessory ecosystem
  • Real-timeNo on-board MCU · needs ESP32 / Pico sidecar for servo timing
  • Power~5–8 W typical, peaks 12 W under load + HAT
  • OSRaspberry Pi OS (Debian) · everything just works
  • EcosystemMassive · ROS2, OpenCV, Hailo SDK, hobby robot codebases everywhere
  • SA stockPiShop ZA stocks everything · same-week delivery
OPTION B · MORE ELEGANT

Arduino UNO Q (4 GB)

SBC + MCU on one PCB · integrated · lower ceiling
  • MPU4× Cortex-A53 @ 2.0 GHz (QRB2210)
  • RAM4 GB LPDDR4X max
  • MCUSTM32U585 Cortex-M33 @ 160 MHz · 786 KB SRAM · on-board, real-time
  • NPUNone · Hexagon DSP available but no NPU class
  • StorageeMMC 32 GB on-board · no NVMe
  • CameraMIPI-CSI · single port
  • GPIO3.3 V on MCU side / 1.8 V on MPU side · MIPI-DSI display out · Qwiic
  • Real-timeBuilt in via STM32 + bridge — no sidecar needed
  • Power~2–4 W typical · noticeably lighter on the LiPo
  • OSDebian + Arduino App Lab on top
  • EcosystemBrand new (announced 2026) · few hexapod codebases yet
  • SA stockNot yet · international order, 4–6 wk + customs

02 · How much LLM can each one actually run locally?

Quantized to Q4 (the practical sweet spot), running via llama.cpp. Numbers are realistic ballparks, not lab maxima. Useful = comfortably above reading speed; painful = waits between words; no = won't load. The AI HAT does not help LLM workloads today — Hailo-8 is optimised for vision/CNN.

Model Use case UNO Q (A53 ×4, 4 GB) Pi 5 8 GB (A76 ×4) Pi 5 16 GB (A76 ×4)
Qwen 2.5 0.5B Q4 Intent parsing · classifier 10–15 tok/s 25–35 tok/s 25–35 tok/s
TinyLlama 1.1B Q4 Short replies · routing 4–6 tok/s 15–20 tok/s 15–20 tok/s
Phi-3-mini 3.8B Q4 Persona · brief reasoning ~1–2 tok/s 6–9 tok/s 6–9 tok/s
Llama 3.2 3B Q4 Persona · tool calling ~1–2 tok/s 5–8 tok/s 5–8 tok/s
Mistral 7B Q4 Real reasoning · multi-turn won't load 2–3 tok/s 3–4 tok/s
Llama 3 8B Q4 Genuine assistant quality won't load tight · 1 tok/s ~2 tok/s
Llama 3 13B Q3 Above hobby scale no no ~1 tok/s, marginal

Translation: UNO Q tops out at tiny intent/classifier models. Pi 5 8 GB can comfortably run 3–4B persona models. Pi 5 16 GB unlocks 7–8B reasoning models if you really want them on-board. None of this beats cloud — but it changes what's possible offline.

03 · Vision (where the AI HAT actually earns its keep)

Workload UNO Q (CPU) Pi 5 (CPU only) Pi 5 + AI HAT+ (26 TOPS)
YOLOv8 nano (320 px) ~8–12 FPS ~25–35 FPS 120+ FPS
YOLOv8 small (640 px) ~2–4 FPS ~8–10 FPS 60+ FPS
MediaPipe pose / hands real-time real-time real-time (CPU)
Whisper tiny (STT) ~1× real-time ~3× real-time ~3× real-time

For a hexapod doing obstacle avoidance, room recognition, and "is the cat there", Pi 5 alone is enough. The AI HAT becomes worth it when you want larger detection models running continuously without burning CPU budget.

04 · For your house-creature, specifically

+ PI 5 + AI HAT WINS WHEN…
  • You want local LLM at persona quality 3B+ models for the "voice" without paying cloud round-trip latency or API costs.
  • Heavy continuous vision YOLO running every frame, multi-camera, room mapping. The HAT pulls the load off the CPU.
  • Future-proof ceiling 16 GB RAM means you can swap in bigger models for years without rebuilding.
  • Available now, lots of community PiShop ZA has everything in stock. Every hobby robot tutorial assumes Pi.
  • Tinkerable Standard PCIe, GPIO, USB 3 — bolt anything onto it later.
− …BUT THE COST IS
  • Two boards, not one Pi 5 still needs an MCU sidecar (ESP32 / Pico) for the 50 Hz servo loop. More wiring, two firmware projects.
  • More power draw 5–8 W typical vs UNO Q's 2–4 W. Roughly halves on-LiPo runtime — needs a bigger BEC and a runtime check.
  • Bigger physical footprint Pi 5 + active cooler + HAT stack is a tall sandwich. UNO Q is one flat board.
  • ~2–3× the BOM cost See the cost section below.
  • Heat Active cooler mandatory under sustained load. Worth noting in an enclosed chassis.

05 · Cost (SA, landed, rough)

OPTION A · PI 5 STACK

Full LLM + vision build

ItemPrice (ZAR)
Raspberry Pi 5 (16 GB)R3 200
Active coolerR220
AI HAT+ (26 TOPS Hailo-8)R2 500
NVMe HATR500
NVMe SSD 256 GBR600
Camera Module 3R900
ESP32 sidecar (you already have this)R150
PSU + cables + microSDR450
Total~R8 520

Trim-down: Pi 5 8 GB + no AI HAT + no NVMe = ~R3 800. Still beats UNO Q on LLM if cloud isn't enough.

OPTION B · UNO Q STACK

Integrated build

ItemPrice (ZAR)
Arduino UNO Q (4 GB / 32 GB)~R1 800
International shipping + customs~R400
MIPI camera module~R600
PSU + cablesR250
Total~R3 050

Cheaper, less to wire. But local LLM ceiling is tiny — you lean harder on cloud Claude/GPT.

06 · Recommendation

DECISION

If you want local LLM and lots of headroom: Raspberry Pi 5 16 GB + AI HAT+ + ESP32 sidecar. You get a 3–8B-parameter local model for the persona, real NPU horsepower for vision, every robot library that exists, and you can buy it from PiShop tomorrow.

If you want integration and elegance, and you're OK leaning on cloud LLM for the voice: Arduino UNO Q. Half the cost, half the power, real-time MCU built in, one board.

For your specific goal — "walks the house, texts me when it sees something": The persona is asynchronous (no latency pressure). Cloud Claude/GPT does the writing for cents per day. Both platforms are fine for vision at the scale you need. UNO Q is the more elegant fit unless you specifically want offline LLM or a Pi-grade vision pipeline. But Pi 5 is the path of least surprise — every problem has been solved by someone before you.

My pick if it were my robot: Pi 5 16 GB + active cooler + Camera Module 3 + ESP32 sidecar, no AI HAT for the first build. Add the HAT later if vision actually hits a wall. That's ~R4 000 instead of R8 500, and it preserves the upgrade path.