class 18 · unit II wrap · multi-device architectures, devices and workloads
Describe the job before you pick the chip.
Eight classes of sensors and processors. Today we turn any product into five numbers, match those numbers to hardware, and split the work across devices when one is not enough. Then the unit quiz.
Theory, recap and quiz run to slide 13; the selection exercise is one block at the end (slides 14 to 16). Print four product cards per team.
last class's exit ticket
Pick the hardware, with one reason.
product
choice
the reason that decides it
10 million fitness bands
ASIC
volume pays the design cost; battery life is the product
research radar, processing changes monthly
FPGA
rewire after deployment; low volume
40 CCTV cameras in a mall
edge server
many streams, uplink bandwidth, video stays in the building
Every good answer used a few facts about the job: how many, how often, how fast, on what power. Today we make that a habit.
Two minutes. Note who argued from numbers rather than from the name of the chip.
try it live · pick a workload
Five numbers describe any workload.
The doorbell row is the plan's worked example. Figures are orders of magnitude for teaching, not a specific product's measurements. The bars for compute and memory are on a log scale: each step is ten times more.
Workloads come in four shapes. Each points to a different device.
Always-on: tiny but never stops, so standby power dominates (wake-word chips, sensor hubs). Periodic: mostly asleep, so sleep current dominates (MCU in deep sleep). Bursty: wake fast, finish fast, sleep (MCU, maybe with an NPU). Continuous heavy: throughput and cooling dominate (edge server).
try it live · a battery doorbell that wakes on the PIR
A bursty device lives or dies by its sleep current.
visitors per day
awake per visitor
sleep current
energy awake
energy asleep
Assumes 0.3 W awake (Class 17's estimate) and three AA cells, about 11 Wh. The ESP32-S3 datasheet gives deep-sleep currents of a few to tens of microamps depending on what stays powered; the dev board's power LED and USB bridge draw milliamps, which is why products use their own boards. Click the careless board: about 11 years becomes about two weeks. Illustrative, but the lesson is real: a careless sleep current wipes out everything else.
The Unit II menu, matched to the four shapes.
shape
usual first choice
when to move up
always-on
a low-power core, sensor hub or audio chip
millions of units: a custom ASIC (Class 17)
periodic
an MCU in deep sleep (Class 15)
rarely; this is the MCU's home ground
bursty
an MCU with vector instructions
model too big: MCU + micro-NPU or SoC + accelerator (Class 16)
continuous heavy
SoC with an NPU, or a GPU box
many streams in one place: an edge server (Class 17)
And the exceptions: hard deadlines or odd interfaces point to an FPGA, whatever the shape.
"Cheapest thing first" from Unit I: start at the top of the middle column and move right only when a number forces you to.
try it live · click an architecture
When one device is not enough, split the work.
These are Class 5's patterns again (cascade, gateway, hierarchy), now with a chosen piece of hardware at each tier. The rule in every case: the cheapest, lowest-power tier filters, so the expensive tier only wakes for what matters.
Your ESP32-S3 is already two devices.
main cores · 2 × LX7
240 MHz, vector instructions, Wi-Fi. Powerful and hungry. Asleep most of the time.
ultra-low-power coprocessor
A tiny core that keeps running in deep sleep: it can read a sensor, check a threshold and wake the main cores only when something changes.
That is "sensor + hub" on one chip: the same idea as the always-on sensor hub in every phone.
The S3 has a ULP-RISC-V and a ULP-FSM coprocessor; programming them needs ESP-IDF rather than plain Arduino sketches, so we name it and move on. Unit V can return to it for projects that need long battery life.
unit recap · our running example, with real hardware
Eight classes, one doorbell.
Reveal one part at a time and ask students to name the class and one fact from it before you click. The camera stays on the diagram even though our kit lab used a webcam.
Unit II on one page.
class
topic
the sentence to remember
10
Sensors, signals, sources
Every sensor is a chain: transducer, conditioning, ADC, interface. Range and resolution trade off.
11
Acoustic and vibration
Sample above twice the highest frequency; MEMS mics on I2S, accelerometers on I2C.
12
Visual and scene
A VGA RGB565 frame is 600 KB: only PSRAM holds it. Thermal, depth and event cameras see differently.
13
Motion, position, force, touch
Accelerometers give tilt, gyros drift; fuse them. Capacitive touch is nearly free on the S3.
14
Optical, EM, environmental, bio, chem
PIR wakes, radar sees through the dark, gas sensors drift and cross-react.
15
Architecture, MCUs, DSPs
Moving data costs more than maths. SIMD and int8 do more per instruction.
16
SoCs and accelerators
NPUs win by reuse and low precision. Read TOPS with four questions; prefer TOPS/W.
17
FPGAs, ASICs, edge servers
Flexible to efficient. Break-even = NRE ÷ unit saving. Edge servers serve many streams on site.
Revision sheet for the CIE: tell students to photograph it. Ask a student to say each sentence before it appears.
unit II quiz · vote with hands, then click
Questions 1 to 10 of the Unit II question bank, as multiple choice. For a written CIE, use the bank's open versions. The score counts first clicks only.
where we are
Data is coming in. Hardware is chosen. Now the model.
III
feature engineering
turning streams into numbers a model can use
III
working with streams
windows, buffers and real-time constraints
III
combining sensors
fusion: more than one view of the world
III
feasibility
should we build this at all?
One minute. Then the selection exercise.
selection exercise · teams of four · four product cards
Five numbers, then a choice, for each product.
13:00
beehive monitor
weight, temperature and hive sound in a field; battery and solar; must last a season; alert if the colony is about to swarm
wearable fall detector
worn by an elderly person all day; detect a fall within a second; battery should last a week; call for help
factory with 200 motors
vibration on every motor; flag bearing wear days early; mains power; one control room
railway station, 50 cameras
count crowds and spot unattended bags; mains power; video must not leave the site; alerts within seconds
The menu: MCU · MCU + micro-NPU · SoC · SoC + accelerator · FPGA · ASIC · edge server · or a combination. One sentence of justification per product, using the five numbers. Two teams will present; the class challenges.
Circulate. The fall detector is the trap: teams often reach for a camera or a big model. Push them to the five numbers: always-on, tiny compute, a week on a battery.
click a product after two teams have presented
Defensible answers, not the only ones. Reward any choice whose justification uses the five numbers.
selection exercise · worksheet
One box per product.
beehive monitor
duty cycle · compute · memory · latency · power → choice, and one sentence why
wearable fall detector
duty cycle · compute · memory · latency · power → choice, and one sentence why
factory with 200 motors
duty cycle · compute · memory · latency · power → choice, and one sentence why
railway station, 50 cameras
duty cycle · compute · memory · latency · power → choice, and one sentence why
Collect these: they make a good CIE component, and they reuse question 18 of the bank.
exit ticket · on paper, before you leave
Describe one product you use every day as a workload in the five numbers, and name the device class you think runs it.
That is Unit II. Next: Unit III, models, tools and feasibility, starting with feature engineering on data streams.
Good answers commit to rough numbers (orders of magnitude) rather than leaving any blank.