E.A.5 Edge Device Deployment

Edge deployment means the model runs near the user, camera, machine, or sensor. The main problem is not model accuracy first; it is whether the device can run the system reliably for a long time.

What You Need

• Python 3.10+
• No external packages
• A target scenario, such as camera classification, factory inspection, or offline form reading

The Four Checks

• Memory: model, runtime, input buffer, and service all need RAM.
• Power: a device that can run once may still overheat or throttle.
• Latency: some tasks need instant response; some can wait.
• Offline mode: if the network is unstable, the device still needs a local fallback.

Run A Compatibility Filter

Create edge_fit.py:

devices = [
    {"name": "edge-a", "memory_mb": 512, "power_w": 8, "offline": True},
    {"name": "edge-b", "memory_mb": 2048, "power_w": 15, "offline": False},
    {"name": "edge-c", "memory_mb": 4096, "power_w": 25, "offline": True},
]

model = {
    "name": "int8-small-classifier",
    "memory_mb": 700,
    "power_w": 10,
    "latency_ms": 65,
    "requires_offline": True,
}

for device in devices:
    reasons = []

    if device["memory_mb"] < model["memory_mb"]:
        reasons.append("memory")
    if device["power_w"] < model["power_w"]:
        reasons.append("power")
    if model["requires_offline"] and not device["offline"]:
        reasons.append("offline")

    status = "FIT" if not reasons else "CHECK " + ",".join(reasons)
    print(device["name"], status)

Run it:

python edge_fit.py

Expected output:

edge-a CHECK memory,power
edge-b CHECK offline
edge-c FIT

Read the result from left to right: edge-c is not automatically the fastest or cheapest device, but it is the only one that satisfies the deployment constraints.

Make It More Real

Change model["memory_mb"] from 700 to 350 and run again. edge-a still fails because power is too low. This shows why edge deployment is a multi-constraint problem.

Practical Edge Checklist

Before calling a device “ready,” verify:

1. It can start from cold boot.
2. It can run for at least 30 minutes without memory growth.
3. It handles network loss.
4. It saves enough logs for remote troubleshooting.
5. It has a simple rollback or replacement path.

Common Mistakes

• Picking the model first, then trying to force it onto a small device.
• Testing only one inference instead of a long-running loop.
• Assuming the device is always online.
• Forgetting that logs, caches, and input images also consume memory.

Practice

Add price_usd to each device and choose the cheapest device that passes all checks. Then add a second model and compare which device works for both.