Two Budgets That Decide Cloud Versus Edge
Ada re-derives this chapter’s own numbers step by step, at full precision
ADA · CALCULATION AUDIT
Two Budgets That Decide Cloud Versus Edge
The chapter rests on two sizing claims: local inference shrinks a 2 Mbps uplink “more than 10,000 times” to about 20 bytes/s, and INT8 quantization takes a 5-million-weight model from 20 MB to 5 MB so it fits a 16 MB gateway. Both decide the cloud-versus-edge split. This audit carries each from the stated inputs to show they are the two budgets that decide cloud versus edge.
Companion to the chapter Edge AI Fundamentals: Why and When — every number here comes from that chapter.
Ada: This deep dive rests on two sizing claims: local inference shrinks an uplink “more than 10,000 times,” and INT8 quantization takes a 5-million-weight model from 20 MB to 5 MB so it fits a 16 MB gateway. Both decide the cloud-versus-edge split, so let me carry each from the stated inputs.
- Raw uplink: a
2 Mbpsstream is2000000 / 8 = 250000 bytes/s = 250 kB/s. - Event uplink: a
200-byte report every10 sis200 / 10 = 20 bytes/s = 160 bits/s. - Reduction:
2000000 / 160 = 12500xsmaller – consistent with “more than 10000 times.” - FP32 model:
5,000,000 weights x 4 bytes = 20,000,000 bytes = 20 MB. - INT8 model:
5,000,000 x 1 byte = 5 MB, a20 / 5 = 4xshrink. - Gateway fit: the INT8 artifact plus a
4 MBtensor arena and2 MBinput buffer is5 + 4 + 2 = 11 MB, leaving16 - 11 = 5 MBheadroom; the20 MBFP32 version overflows the16 MBbudget before any buffers.
The audit conclusion is that quantization is what makes the whole placement legal: at FP32 the model does not fit, so “run it at the edge” is not an option; at INT8 it fits with 5 MB to spare. The design meaning is that both budgets point the same way – the model must be small enough to load and the traffic small enough to send – and a 4x weight shrink plus a 12,500x traffic shrink is precisely what turns a cloud-only inspection stream into a device that reports 20 bytes a second and survives a WAN outage.
Every number above is taken from the chapter’s own material and re-derived step by step.