← Back to the chapter

Technical detail · Latency Budgets for Tier Choice

Separate model speed from whole-response time

The saw-stop teaching model uses a 100 KB request, a 10 Mbps link, complexity 4, no congestion, a 12 ms cloud round trip and a 100 ms target. These are illustrative inputs, not a measured saw-stop deadline.

Request transmission is 100 × 8 ÷ 10 = 80 ms. The response estimate is 80 × 0.12 = 9.6 ms; the model’s floor gives max(0.08, 9.6) = 9.6 ms. Cloud request/network time is (12 ÷ 2 + 80) × 1 = 86 ms. Waiting is 5 + 0 × 1.1 = 5 ms, so no congestion still leaves base waiting time.

Cloud processing is 4 × 1.4 = 5.6 ms. The return congestion multiplier is min(1, 3) = 1; return time is 12 ÷ 2 + 9.6 × 1 = 15.6 ms. Adding the stages gives 86 + 5 = 91; 91 + 5.6 = 96.6; 96.6 + 15.6 = 112.2 ms. The native card rounds to 112 ms, above the 100 ms target.

Edge processing is 4 × 10.5 = 42 ms. Cloud computes faster at 5.6 ms, yet its request path alone takes 86 ms. The model covers transfer through return; sensing and brake motion still belong in the complete physical-response budget. Choose from the whole path rather than inference speed alone.

← Back to the chapter