Google Coral / Edge TPU

1. Buy a Google Coral USB Accelerator ($60) or Coral Dev Board Mini ($100). The Coral Edge TPU is a dedicated INT8 inference chip — 4 TOPS at 2W.
2. pip install pycoral → download a pre-compiled TensorFlow Lite model from coral.ai/models (classification, detection, segmentation, pose estimation).
3. Python script for object detection:

from pycoral.utils import edgetpu
from pycoral.adapters import common
import cv2

interpreter = edgetpu.make_interpreter("mobilenet_ssd_v2_coco_quant_edgetpu.tflite")
interpreter.allocate_tensors()
img = cv2.imread("photo.jpg")
common.set_input(interpreter, img)
interpreter.invoke()
detections = common.get_output(interpreter)
for d in detections:
    print(f"{d.id}: {d.score:.2f} at ({d.bbox.xmin}, {d.bbox.ymin})")

1. First detection in <10 ms — the Edge TPU runs at 100+ fps for MobileNet-class models.
2. Critical limitation: Coral TPU only runs TensorFlow Lite INT8 models that are compiled specifically for the Edge TPU. You can't run PyTorch models, LLMs, or mixed-precision models. Coral is for vision classification/detection/segmentation, not general-purpose AI.
3. For custom models: train in TensorFlow → quantize to INT8 → compile with Edge TPU Compiler → deploy. The compiler is strict — unsupported ops (many common ones) fail.

Google Coral USB Accelerator (~$60) + Raspberry Pi 4 ($55) + Pi Camera ($25) + SD card ($10). Total: ~$150. Runs MobileNet SSD at 100+ fps, PoseNet at 30+ fps, face detection at 100+ fps. For a DIY security camera with person detection: $150 handles real-time detection with <10ms inference latency at 2W power. For a wildlife camera classifying animals: Coral + Pi + solar panel = a $200 self-powered AI camera that runs for months. Coral is the cheapest way to add GPU-class vision inference to a Raspberry Pi. The tradeoff: only TensorFlow Lite vision models.

Coral doesn't scale to "serious" — it's deliberately capped. A Coral Dev Board Micro ($100) maxes out at 4 TOPS INT8, ~300 fps for MobileNet. For production edge vision (factory QA, retail analytics): deploy multiple Coral devices per camera, or upgrade to Jetson Orin Nano ($500, see /hardware/jetson-ai) for 40 TOPS + general-purpose CUDA. Coral's niche is ultra-low-power, ultra-low-cost, single-purpose vision inference. For general-purpose edge AI (LLMs, multi-model pipelines, custom architectures), Coral is the wrong tool. Use Coral when you need: <5W, <$100, vision-only, supported TF Lite ops. Use Jetson for everything else.

The mistake: Buying a Coral TPU, installing PyTorch, and trying to run a Llama model on it — getting "unsupported operation" errors for every layer. Why it fails: The Edge TPU is not a GPU. It's an ASIC that executes exactly one format: TensorFlow Lite INT8 models with a limited op set. No PyTorch. No FP16. No attention mechanisms. No dynamic shapes. It's designed for convolutional vision models (MobileNet, EfficientNet). The fix: Check the Edge TPU supported ops list before designing your model (coral.ai/docs/edgetpu/models-intro). If your model uses any unsupported op (LSTM, attention, custom layers, most normalization ops), the compiler rejects it. Coral is for "camera sees something → classify/detect" workflows. For anything involving transformers, LLMs, or generative models, use a Jetson or x86 edge PC. Coral is a one-trick pony — but for that one trick (real-time vision classification), it's the best pony on the market at 2W.