Appearance
10.2 — Anatomy of a Smartphone
A modern phone contains perhaps twenty billion transistors, forty sensors, six radios and a battery holding more energy than a hand grenade, in a sealed slab seven millimetres thick that survives being dropped and costs less than a bicycle.
Every chapter of this volume is inside it. This chapter takes it apart and points at each one.
1. The system on chip
The main processor, and it is not one processor.
CPU cores, in clusters. A typical arrangement is two large cores at 3 GHz for bursts of demanding work, three medium cores at 2.4 GHz, and three small ones at 1.8 GHz for background tasks.
Why the asymmetry. From Chapter 2.4, dynamic power is \alpha CV^2f, and a high-performance core is physically larger with more capacitance and needs a higher voltage to run fast. A big core doing a small job wastes most of the energy. The scheduler moves each task to the smallest core that can meet its deadline, and the energy saving is a factor of three to five on typical workloads.
GPU — hundreds of small parallel cores, for graphics and increasingly for general parallel computation.
Neural processing unit — a matrix multiplication engine, doing tens of trillions of operations per second at a fraction of the CPU's energy per operation. It exists because the arithmetic is regular and predictable, which is exactly the argument Chapter 3.6 made for FPGAs.
Image signal processor — the camera pipeline of Chapter 5.5, in dedicated hardware. Doing it in software would take seconds per photograph and drain the battery.
Modem — often a separate chip, containing its own DSP cores running the entire physical layer of Chapter 8.1.
Memory controller, display controller, video codecs, security enclave — each a fixed-function block, because each does one regular job constantly and dedicated silicon does it at a tenth the energy.
The pattern worth naming: a modern SoC is mostly accelerators, and the CPU is the part that handles whatever has no accelerator. That inversion happened around 2015 and it is the single biggest architectural change in mobile computing.
Process and packaging
3 nm and 4 nm processes, where the number is a marketing label rather than a physical dimension — actual gate lengths are around 20 nm and the number describes a density generation.
Memory is stacked on top of the SoC — package-on-package — putting the DRAM directly above the processor. Short connections mean lower capacitance, so less energy per bit transferred and higher bandwidth, from the same CV^2f argument.
Thermal management is the binding constraint. A phone has no fan and no heatsink, so sustained power is limited to about 3 to 5 W by how fast the case can shed heat.
Which is why phones throttle. A benchmark that runs for thirty seconds sees full performance; one that runs for ten minutes sees perhaps 60% of it. The peak figure describes a burst, not a capability, and the difference is thermal.
The single most important thing this diagram shows is what is inside the box versus outside it. The CPU, graphics, neural accelerator, image processor, video codec, memory controller and security engine are all on one piece of silicon. That matters for two reasons. Moving data between chips costs far more energy than moving it within one — driving a signal off-chip means charging centimetres of track instead of micrometres — so integration is mostly an energy decision rather than a size one. And a shared memory controller lets the camera pipeline hand a frame to the neural accelerator without copying it anywhere, which is what makes computational photography fast enough to run between shutter presses.
What stays outside, and why. Memory and flash are made on different processes optimised for storage density rather than logic speed, so they remain separate dies even when stacked in the same package. The radios need analogue circuitry and filters that do not shrink with the digital process, and they must sit near their antennas. And the battery, display and camera sensors are physical objects with sizes set by chemistry and optics, not by lithography.
The next sections take the two most demanding of these outside blocks in turn.
2. The radios
Six or more, and they must coexist in a space smaller than a hand.
| Radio | Bands | Chapter |
|---|---|---|
| Cellular | 30–60 bands, 600 MHz – 40 GHz | 8.1 |
| Wi-Fi | 2.4 / 5 / 6 GHz | 8.3 |
| Bluetooth | 2.4 GHz | 8.3 |
| GNSS | 1.2 / 1.5 GHz | 8.2 |
| NFC | 13.56 MHz | 8.4 |
| UWB | 6–8 GHz | — |
The antenna problem is genuinely severe. Chapter 7.6 established that an efficient antenna is about a quarter wavelength. At 700 MHz that is 107 mm — longer than the phone.
The solutions, all of them compromises:
- The metal frame is the antenna, split into segments by insulating gaps. Those plastic strips in the metal band are not decoration — they define the antenna segments.
- Meandering and folding the radiating element to fit its electrical length into a smaller physical space.
- Loading with capacitance to shorten the resonant length, at the cost of bandwidth and efficiency.
- Accepting poor efficiency. A phone antenna at 700 MHz might be 30 to 40% efficient — a loss of 4 to 5 dB compared with a proper antenna, simply accepted.
And the hand is part of the antenna. Holding a phone changes its resonant frequency and absorbs energy, typically costing 3 to 10 dB. Phones use multiple antennas and switch to whichever is least affected, which is what the "antenna gate" episodes of the early 2010s were about — a design where the only antenna could be bridged by a finger.
Coexistence. Wi-Fi and Bluetooth share 2.4 GHz; cellular bands sit near GNSS frequencies. The chipset coordinates them in time, and filters keep each radio's transmissions out of the others' receivers. A phone contains twenty or more filters and switches in the front-end module, and that module is often the most expensive component after the SoC and the display.
3. The camera
The sensor is a CMOS array, typically 12 to 50 megapixels, with each pixel a photodiode (Chapter 2.2) and its own amplifier.
Pixel size is the fundamental trade. A 1 µm pixel collects a quarter of the light of a 2 µm pixel. More megapixels in the same sensor area means worse low-light performance, which is why the megapixel race stopped and reversed.
Pixel binning is the resolution. A 48-megapixel sensor with 0.8 µm pixels combines groups of four into one 1.6 µm effective pixel, giving a 12-megapixel image with four times the light per pixel. Full resolution in daylight, binned in low light, and the sensor is designed for both.
The Bayer filter puts a red, green or blue filter over each pixel in a pattern with twice as many green as red or blue — because the eye's luminance sensitivity peaks in the green, exactly as Chapter 5.5's YCbCr weights showed.
Demosaicing reconstructs full colour at every pixel by interpolating from the neighbours, and it is the first step of the image pipeline.
The pipeline, all in dedicated hardware:
- Black level and defect correction
- Demosaicing
- White balance
- Noise reduction
- Lens distortion and vignetting correction
- Sharpening
- Tone mapping
- Encoding to JPEG or HEIC
Computational photography is what actually distinguishes modern phone cameras, and it is where the interesting engineering is:
HDR captures several exposures and merges them, extending the dynamic range far beyond what one exposure can hold.
Night mode captures many frames over several seconds, aligns them to cancel hand movement, and averages. Averaging N frames reduces random noise by \sqrt N — the same result as Chapter 5.3's Welch averaging, applied to images. Sixteen frames gives four times the signal-to-noise ratio, which is worth two stops of exposure.
Portrait mode estimates depth — from two cameras, from a dedicated depth sensor, or from a neural network — and blurs the background accordingly.
The honest summary: a phone's optics are far worse than a camera's, with a lens a few millimetres across and a sensor the size of a fingernail. The images are comparable because the processing compensates, and the processing is doing more work than the optics.
Optical image stabilisation moves the lens or sensor to counter hand shake, driven by the gyroscope of Chapter 10.1 and voice coil actuators. It buys three to four stops — the difference between 1/60 s and 1/8 s hand-held.
4. The display
OLED — each pixel emits its own light, so black pixels are genuinely off. Infinite contrast, and power that depends on the image content, which is why dark mode saves battery on OLED and does nothing on LCD.
LCD — a backlight always on, with liquid crystal shutters and colour filters. Cheaper, and its blacks are grey.
Subpixel arrangements. OLED often uses a diamond pattern with fewer blue and red subpixels than green, because blue OLED material degrades fastest and larger blue subpixels can be driven less hard for the same brightness. This is also why burn-in appears as a blue-tinted ghost of static elements.
Refresh rate. 60 Hz was standard; 120 Hz is now common. Variable refresh drops to 10 Hz or lower for static content, saving substantial power — because the display controller is a large fraction of the idle power budget.
Touch sensing is projected capacitance: a grid of transparent electrodes, with a finger changing the capacitance at an intersection.
\Delta C\approx1\ \text{pF against a background of } 10\ \text{pF}
A 10% change, detected 120 times a second across a grid of perhaps 30 by 50 nodes. The noise problem is severe — the display's own driving signals couple into the touch grid, so touch sensing is synchronised to happen between display refreshes.
5. Power
Battery: a single lithium-polymer cell, 3.7 to 3.85 V nominal, 3000 to 5000 mAh, so 11 to 19 Wh.
The power management IC contains a dozen or more switching converters (Chapter 9.5) and low-dropout regulators, generating perhaps twenty different supply rails: 0.6 V for the CPU cores, 1.1 V for memory, 1.8 V for peripherals, 3.3 V for sensors, and high voltages for the display and camera actuators.
Dynamic voltage and frequency scaling. Each block's voltage and clock are adjusted continuously to the minimum that meets the current demand. Because power goes as V^2f, this is worth a factor of several.
Where the power goes in typical use:
| Function | Power |
|---|---|
| Display at medium brightness | 500–1500 mW |
| SoC under load | 2000–5000 mW |
| SoC idle | 50–200 mW |
| Cellular modem, active data | 500–2000 mW |
| Modem idle, registered | 5–20 mW |
| Wi-Fi active | 300–800 mW |
| GPS | 30–100 mW |
| Standby total | 5–15 mW |
Standby life: a 4000 mAh (15 Wh) battery at 10 mW gives
\frac{15}{0.010}=1500\ \text{hours}=62\ \text{days}
Active use at an average of 1.5 W:
\frac{15}{1.5}=10\ \text{hours}
Which matches reality, and shows that a phone's battery life is set almost entirely by screen-on time.
Charging follows the CC-CV protocol of Chapter 9.3, with negotiation over USB Power Delivery to select a voltage from 5 to 20 V. Fast charging above about 30 W typically splits the battery into two cells charged in parallel, halving the current per cell and therefore the heat.
6. The sensors
Around forty in a modern phone, though many are on the same chip.
| Sensor | Purpose |
|---|---|
| Accelerometer | orientation, step counting, gesture |
| Gyroscope | rotation, stabilisation |
| Magnetometer | compass |
| Barometer | altitude, floor detection, weather |
| Proximity | screen off during a call |
| Ambient light | brightness, white balance |
| Fingerprint | authentication |
| Face depth (dot projector) | authentication |
| Temperature (several) | thermal management |
| Hall effect | case open or closed |
The barometer deserves a note because its use is not obvious. It resolves about 10 cm of altitude change, and its purpose is emergency call location — telling responders which floor of a building you are on, which GPS cannot do (Chapter 8.2's vertical geometry problem).
Sensor fusion runs on a dedicated always-on low-power core, combining accelerometer, gyroscope and magnetometer as Chapter 10.1 described. Running it there rather than on the main processor is what lets step counting continue for days on a fraction of a percent of the battery.
7. Security hardware
A secure enclave — a separate processor with its own memory, its own boot ROM, and no path for the main processor to read its keys.
What it holds: biometric templates, payment keys, and the device encryption key.
The important architectural point: the fingerprint image never leaves the enclave, and the main operating system never sees it. The enclave answers a yes-or-no question, and that answer is what the rest of the system acts on.
Secure boot. Each stage verifies the signature of the next before running it, rooted in a public key burned into the silicon at manufacture. A chain of trust from immutable hardware upward, so modified firmware simply does not run.
Hardware encryption. The storage is encrypted with a key derived from the enclave's key and the user's passcode. Because the key is in hardware, brute-forcing the passcode requires the device itself — and the enclave enforces increasing delays between attempts, which is what makes a six-digit passcode adequate.
8. Assembly
The main board is a rigid-flex printed circuit with 10 to 14 layers, often two boards stacked and connected by a flexible section, folded to fit around the battery.
Components down to 0201 metric — 0.25 by 0.125 mm — and increasingly 01005. Placed by machines at tens of thousands per hour with 20 µm accuracy.
Assembly is almost entirely automated except final inspection and packing.
Sealing. IP68 rating means the phone survives immersion. Achieved with adhesive gaskets, waterproof mesh over the speaker and microphone ports, and conformal coating on the board.
And it is the main reason phones stopped having removable batteries. A user-replaceable battery needs a removable cover with a seal that a user opens repeatedly, and a seal that is opened is a seal that eventually fails. Whether the trade was worth it is a fair argument; the engineering reason is straightforward.
9. Where each chapter of this volume appears
Part 1 — Circuit theory. Every power rail, every filter, every impedance match.
Part 2 — Devices. Twenty billion MOSFETs. Op-amps in the audio path. Diodes protecting every input.
Part 3 — Digital. The processors, the memory, the converters at every analog boundary.
Part 4 — Signals and systems. Every filter, every transform, the sampling theorem at every converter.
Part 5 — DSP. Audio codecs, echo cancellation, the camera pipeline, the modem's equaliser.
Part 6 — Control. The power management loops, the camera's autofocus and stabilisation, the modem's power control, the thermal governor.
Part 7 — Communication. Every radio, every modulation scheme, every antenna.
Part 8 — Wireless. Cellular, GPS, Wi-Fi, Bluetooth, NFC.
Part 9 — Power. The battery, the charger, the converters, the wireless charging coil, the haptic motor.
Part 10 — Embedded. All of it, integrated.
The point of the list is not completeness. It is that a phone is not an application of electronics engineering — it is very nearly all of electronics engineering, in one object, and that is why it took a century of accumulated work to make one.
Chapter 10.3 closes the volume by looking at what happens when devices like this stop being individual objects and start being a network of them.
Every formula above, built from scratch
None of the results in this chapter are worth memorising, because each one can be rebuilt in under a minute from something simpler. What follows is that rebuilding, one result at a time, so the formula and the reason for it sit on the same page as the explanation that needed them.
Smartphone
Dynamic power (Chapter 2.4):
P=\alpha CV^2f
Which is why heterogeneous cores work: a large core has more capacitance and needs a higher voltage, so running a small task on it wastes most of the energy.
Battery life:
t=\frac{E_{battery}}{P_{avg}}
Worked: 15 Wh at 1.5 W average gives 10 hours; at 10 mW standby gives 62 days.
Thermal limit:
P_{sustained}=\frac{T_{max}-T_{ambient}}{R_{th}}
With no fan and no heatsink, a phone's R_{th} to ambient is about 8 °C/W, so a 40 °C rise permits 5 W sustained — and that, not the silicon, is what limits sustained performance.
Pixel binning:
\text{effective pixel area}=n\times\text{single pixel area}, \qquad \text{SNR gain}=\sqrt n
Four 0.8 µm pixels binned give a 1.6 µm effective pixel with twice the SNR.
Touch capacitance:
C=\frac{\varepsilon A}{d}, \qquad \Delta C\approx1\ \text{pF against 10 pF background}
Antenna length:
\ell=\frac{\lambda}{4}=\frac{c}{4f}
At 700 MHz that is 107 mm — longer than the phone, which is why phone antennas are folded, loaded and inefficient.
Display power on OLED is proportional to the average pixel luminance, which is why dark mode saves energy there and not on LCD.
What the next chapter fixes
A phone is one device with mains-scale engineering behind it and a battery you charge daily. Chapter 10.3 covers the opposite constraint: devices that must run for years on a cell you will never replace, where the radio dominates every design decision and the right architecture is decided by arithmetic about energy rather than by preference.