approximate computing
Suppose you are picking a wall color from a photo. Does it matter if one pixel is off by an imperceptible shade? Of course not — your eye cannot tell, and the answer 'pale blue' is just as good. A great deal of modern computing is like this: streaming video, machine learning, sensor data, image processing. The results are consumed by imperfect human senses or are already noisy and statistical, so being exactly, bit-perfectly correct buys nothing. Approximate computing is the deliberate strategy of trading a little accuracy you do not need for large savings in energy, time, or hardware.
The idea is that exactness is expensive, and you should pay for it only where it actually matters. Concretely, approximate computing spans several tricks. You can use reduced precision — computing in 8-bit or 16-bit numbers instead of 32- or 64-bit, which is smaller, faster, and far more energy-efficient, and is exactly why machine-learning hardware leans on low-precision arithmetic. You can build deliberately simplified circuits, such as an adder that is right almost always but occasionally off in the last bit, in exchange for being smaller and cooler. You can skip work — sampling some data instead of all of it, or stopping an iterative computation early once the answer is 'good enough'. In each case you accept a controlled, bounded error and pocket the savings.
The discipline is in the word controlled, and the honest danger is exactly there. Approximate computing only works where errors are tolerable and quantifiable, and it is catastrophic where they are not: you must never approximate a bank balance, a cryptographic key, a memory address, or a program's control flow — a single wrong bit there corrupts everything. The hard engineering question is which parts of a computation can be sloppy and by how much, and bounding that error rigorously is not easy. Used wisely on error-tolerant workloads, approximation is a powerful lever in the post-Moore toolkit; used carelessly, it silently produces wrong answers that look plausible.
A neural network for recognizing photos is retrained to run in 8-bit numbers instead of 32-bit. The accuracy drops a negligible fraction of a percent — invisible in practice — but the chip uses far less memory and energy and runs several times faster. Reduced precision is the most widely deployed form of approximate computing exactly because vision and language models tolerate it well.
Spend exact bits only where they matter; on error-tolerant work, lower precision buys big energy and speed wins.
Approximation is safe only where error is tolerable and bounded. Never approximate addresses, control flow, cryptographic keys, or financial values — one wrong bit there is not a small error, it is a corrupt result that may look perfectly plausible.