Material Insights · Understanding the Material World

Light, semiconductors and information: from the transistor to the network: Bits, noise and networks

All courses / Understanding Engineering / Module 9: Light, semiconductors and information: from the transistor to the network Part 5 of 5

Bits, noise and networks

A chip computes, but value comes from moving results around. This part explains why digital signals survive noise, how Shannon's formula caps the rate of any channel, how a tap on a phone becomes a chain of radio, fibre and servers, and what the data centre at the end of the chain costs in electricity.

5.1 A bit is a choice between two, and sending only two levels lets a receiver wipe out noise at every relay

A bit is the answer to one yes-or-no question; eight bits make a byte. An analogue signal, such as the voltage from a microphone, can take any value, so any noise added in transit is part of the answer. Photocopy a photocopy a few times and the page turns grey and blurred, because each copy adds its own specks to the last. A digital signal sends only two voltages, say 0 and 1 V, and the receiver decides which was meant by comparing with 0.5 V. Noise is harmless unless it crosses the threshold, so the receiver can make a fresh, clean copy to send onwards.

Noise is random with a typical size, and the chance of a wrong decision falls steeply with the threshold's distance in units of that size. If noise is 0.1 V, the threshold is 5 noise-sizes away and an error occurs about 3 times in 10 million bits. Cut the noise to 0.07 V and the odds fall to about 1 in 2 million million. Error-correcting extra bits push this far lower.

To turn sound or pictures into bits we sample them. A signal containing frequencies up to f_max must be sampled at least twice as fast, or the high notes fold back into false low ones. A telephone line passes up to 3.4 kHz, so it is sampled 8,000 times a second at 8 bits, which is 64 kb/s, the old fixed-line channel rate.

Compression drops what the eye or ear will not miss and what the data repeats. A 1920 x 1080 video at 30 frames a second is 1.5 Gb/s raw, and a stream of 4 Mb/s is 370 times smaller. The trade is computing power at each end, and a small loss of quality.

Data Estimate 8: raw bit rates against what is sent
SignalRaw rate: workingAs sentCompression
Telephone voice8,000 samples/s x 8 bits = 64 kb/sabout 12 kb/s on a mobile callabout 5 times
Video call, 720p at 30 frames/s1,280 x 720 x 24 bits x 30 = 664 Mb/sabout 1.5 Mb/sabout 440 times
Phone photo, 12 megapixels12 million x 24 bits = 288 Mbit = 36 MBabout 3 MB as JPEGabout 12 times
HD video, 30 frames/s1,920 x 1,080 x 24 bits x 30 = 1.5 Gb/sabout 4 Mb/s streamedabout 370 times

Raw rates and ratios calculated; the sent figures are typical and vary with codec and quality setting. Noise error odds: Q(5) = 3 x 10^-7 for a threshold 5 standard deviations from each level.

Source: Bloomfield, How Things Work (Wiley, 2015), ch. 14, 'Optics and Electronics', section 'Optical Recording and Communication' (analogue and digital, binary representation); example and arithmetic by the author
  • Bloomfield, How Things Work (Wiley, 2015) · ch. 14, 'Optics and Electronics' (analogue and digital representations, bits and bytes)

5.2 Shannon showed that every noisy channel has a capacity set by its bandwidth and its signal-to-noise ratio

In 1948 Claude Shannon proved that a channel with bandwidth B hertz and a signal-to-noise power ratio S/N has a capacity C = B x log2(1 + S/N) bits a second. Below that rate, a clever enough code can make errors as rare as you like. Above it, no code can. Capacity grows in direct proportion to bandwidth: double the spectrum, double the rate. But it grows only as the logarithm of power: doubling signal power adds about one bit per hertz, not double.

So there are three things to buy. Bandwidth, which is spectrum, rationed and auctioned. Signal-to-noise, which costs battery power, interferes with neighbours and falls with distance. Or parallel paths: antenna arrays create separate spatial streams, each like an extra channel, and fibre uses two polarisations and many wavelengths.

A voice line of 3.1 kHz at 35 dB signal-to-noise (a ratio of 3,162) gives 36 kb/s, and the dial-up modems of the 1990s, which topped out at 33.6 kb/s, were within a whisker of the ceiling. The 56 kb/s modems beat it only because the server end was digital, which removed one noisy conversion. Today's 5G mid-band channels of 100 MHz at 20 dB give 0.67 Gb/s per stream and, with four streams, 2.7 Gb/s. The fibre row repeats the bound from Part 2.

Shannon also gave a floor for energy: even with unlimited bandwidth, each bit needs at least 0.69 times the noise energy per hertz, a minimum of the same kind as the Landauer limit in Part 4.

Data Estimate 9: Shannon capacity for four channels
ChannelBandwidth BSignal-to-noiseBits per second per HzCapacity
Telephone line3.1 kHz35 dB (3,162)11.636 kb/s
4G carrier, one stream20 MHz15 dB (32)5.0100 Mb/s
5G mid-band, one stream100 MHz20 dB (100)6.7670 Mb/s
5G mid-band, four streams100 MHz20 dB (100)26.62.7 Gb/s
Fibre, C-band, two polarisations4.4 THz20 dB (100)13.358 Tb/s

Capacity = B x log2(1 + S/N). 10 dB is a ratio of 10, 20 dB a ratio of 100, 30 dB a ratio of 1,000. These are ceilings; real links reach 50 to 80 per cent of them.

Source: Shannon, C. E., 'A Mathematical Theory of Communication', Bell System Technical Journal 27 (1948), channel capacity of a band-limited channel with noise; arithmetic by the author
  • Shannon, C. E., 'A Mathematical Theory of Communication', Bell System Technical Journal 27 (1948) · channel capacity of a band-limited channel with noise

5.3 A tap on a phone crosses radio, fibre and routers, and the speed of light in glass sets a floor on delay

Follow a tap on a video link. The phone's radio signal reaches a nearby tower. Three sector panels each serve a slice of the horizon, with, say, 100 MHz of channel in the 3.5 GHz band. At an average of 3 bits per second per hertz, each sector gives 300 Mb/s, the site 900 Mb/s, and at 5 Mb/s per video stream, some 180 viewers at once. So capacity per tower is limited, and operators raise it by adding towers whose cells are smaller, which lets the same frequencies be reused a few kilometres away.

From the tower the data goes by fibre, or sometimes a microwave radio hop, to a metro exchange, then through routers to an internet exchange or a cache of a content delivery network close to the viewer. Data travels in packets, each wrapped with an address, and routers forward each one by itself. A packet network shares its links among thousands of conversations, unlike the old telephone exchange that reserved a 64 kb/s circuit for every call, silence included.

Distance has a price in time. Light in fibre moves at 300,000 / 1.468, or 204,000 km a second, or 4.9 microseconds per km. A cache in the same city answers in under a millisecond. The round trips in the note are minimums; queues and detours add more. That is why streaming firms and banks put servers near their users.

Process flowchart A tap on a phone becomes a video from a server
Inputs Tap on a videoa request of a few hundred bytes Process steps 1 Phone and tower

radio at 1.8 to 3.5 GHz; a sector serves about 300 Mb/s shared

2 Backhaul to metro exchange

fibre, or a microwave radio hop; about 5 microseconds per km

3 Routers and internet exchange

packets forwarded one by one; a cache may answer here

4 Core and long-haul fibre

submarine or terrestrial; 0.2 dB per km, amplified every 50 to 100 km

5 Data-centre server

finds, encodes and sends the video

Outputs Video back on the screenabout 4 Mb/s for HD

Minimum round-trip time (light in fibre, no queues): 50 km, 0.5 ms; Mumbai to Chennai on an assumed 1,500 km route, 15 ms; Mumbai to Singapore on an assumed 5,500 km route, 54 ms; one geostationary satellite hop each way, 480 ms.

Source: Bloomfield, How Things Work (Wiley, 2015), ch. 12, 'Electromagnetic Waves' (bandwidth); ch. 14, 'Optics and Electronics' (optical communication); arithmetic by the author
  • Bloomfield, How Things Work (Wiley, 2015) · ch. 12, 'Electromagnetic Waves' (radio, bandwidth); ch. 14, 'Optics and Electronics' (optical communication)

5.4 A data centre turns nearly all its electricity into heat, so the cost of information is power in and heat out

A data centre is a hall of racked servers (see the product 'Data-centre server'). Almost every watt that enters a processor leaves as heat, so the building must remove it as fast as it arrives, through fans, chilled water or liquid loops. The ratio of total power to the power reaching the IT equipment is the power usage effectiveness, PUE; 1.4 means 0.4 W of cooling and power conversion for each 1 W of computing.

The table sizes a mid-size facility of 10 MW of IT capacity. Run at 70 per cent on average, with a PUE of 1.4, it uses 9.8 MW and some 86 GWh a year. At ₹7 per kWh, that is ₹60 crore. That equals the yearly use of about 70,000 homes at 100 kWh a month. The heat leaves partly through the evaporation of water: removing 9.8 MW with the latent heat of water, 2.26 MJ/kg, takes 4.3 kg a second, about 375 tonnes a day. Cooling by air or by closed liquid loops saves water but uses more electricity. In a water-short region that trade decides the site.

Globally, the IEA reported that data centres used about 485 TWh in 2025, up 17 per cent in a year, and expects close to 950 TWh by 2030, about 3 per cent of world electricity, with the electricity use of AI-focused centres expected to triple. The chips of Part 4 are why: each does more per joule every year, but demand grows faster.

Data Estimate 10: a 10 MW data centre, and an hour of streaming
StepWorkingResult
Average IT load10 MW x 70 per cent7 MW
Total with cooling, PUE 1.47 MW x 1.49.8 MW
Energy a year9.8 MW x 8,760 habout 86 GWh
Electricity bill at ₹7/kWh (assumed)86 million kWh x ₹7about ₹60 crore
Equal to homes using 100 kWh a month86 million / 1,200 kWhabout 70,000 homes
Water if all heat is evaporated9.8 MW / 2.26 MJ per kg4.3 kg/s, 375 tonnes a day
One hour of video on a phone: devices3 W x 1 h0.003 kWh
The same, network share (Kamiya, IEA data, 2019)0.018 kWh per hour0.018 kWh
The same, data-centre share (Kamiya, IEA data, 2019)0.004 kWh per hour0.004 kWh
10 crore viewers, 1 hour a day, a year0.025 kWh x 10^8 x 365about 0.9 TWh
Share of India's 1,850 TWh (2025-26)0.9 / 1,850about 0.05 per cent

All values calculated; phone power and facility figures are assumptions. A 55 inch television uses about 100 W, so the same hour costs 0.1 kWh at home; the 2019 estimate of 0.077 kWh per hour (Kamiya, using IEA data) is an average of such mixes, with devices taking 72 per cent. Network energy per hour has fallen since 2019.

Source: Carbon Brief, George Kamiya, 'Factcheck: What is the carbon footprint of streaming video on Netflix?' (25 February 2020, updated 25 November 2020, read November 2026), based on IEA data, 0.077 kWh per hour of streaming in 2019: devices 72 per cent, networks 23 per cent, data centres 5 per cent; https://carbonbrief.org/factcheck-what-is-the-carbon-footprint-of-streaming-video-on-netflix/
  • International Energy Agency, 'Energy and AI' and 'Key questions on energy and AI' (2025 to 2026, read November 2026) · data-centre electricity use of about 485 TWh in 2025 and about 950 TWh projected for 2030; https://iea.org/reports/key-questions-on-energy-and-ai/executive-summary

See it in the table

H₂O WaterDC Data centresEl Electricity

Check your understanding

Why can a digital signal be relayed many times without building up noise, while an analogue one cannot?

A receiver compares the signal with a threshold and regenerates it, so noise smaller than half the level gap vanishes at each stage. Analogue noise stays in the signal.

A channel has 100 MHz of bandwidth and a signal-to-noise ratio of 100. Shannon's formula gives a ceiling closest to:

C = 100 MHz x log2(101), and log2(101) is about 6.7, so the ceiling is about 670 Mb/s.

A data centre has an IT load of 7 MW on average and a PUE of 1.4. What is its total power draw?

PUE is total power divided by IT power, so total = 7 MW x 1.4 = 9.8 MW.
← Part 4: The transistor and the integrated circuitMark read · Wrap-up →