Amdahl's law calculator

Amdahl's law names the ceiling that leftover serial work puts on parallel speedup. If a fraction P of a program can run concurrently and the rest must stay sequential, adding processors never multiplies runtime by N. The theoretical speedup is one over one minus P plus P over N: the serial remainder stays in the denominator no matter how large N becomes. This Amdahl's law calculator accepts the parallelizable fraction and the processor count, then returns the speedup factor, the serial fraction, parallel efficiency, and the asymptotic serial limit one over one minus P whenever any serial work remains. The arithmetic is a closed-form identity, fully deterministic, with no simulation and no language model. Use it to size a cluster, argue against a naive N-times claim, or check a textbook speedup curve before you buy hardware.

What Amdahl's law measures and how the speedup factor is derived

Amdahl's law is Gene Amdahl's 1967 observation that leftover serial work bounds how much faster a whole program can run, no matter how many processors you apply to the parallel part. Write P for the fraction of single-processor wall-clock time that can run concurrently, and N for the processor count. The sequential fraction is one minus P.

After you divide the parallel work N ways, relative runtime becomes one minus P plus P over N, so the theoretical speedup S of N is the reciprocal of that sum: one divided by one minus P plus P over N. When P equals one, every cycle parallelizes and speedup equals N. When P equals zero, the job is purely serial and speedup stays one for every N.

Between those extremes, speedup grows with N but flattens toward the serial limit one over one minus P. That ceiling is the whole point of the law: a program that is only five percent serial can never exceed a twenty-times speedup, even on a million cores, because the serial five percent remains after the parallel ocean has been sliced infinitely thin. Parallel efficiency is speedup divided by N; it falls as N grows unless P is exactly one.

This calculator evaluates those identities directly. It is not a profiler, not a discrete-event simulator, and not Gustafson's scaled-speedup model, which asks a different question about growing the problem with the machine. If you know P from measurement or from a back-of-the-envelope split of hot loops versus setup, Amdahl's formula gives the theoretical speedup you should expect before cache effects, lock contention, or NUMA traffic make the real curve worse.

Where an Amdahl's law calculator earns its keep

Use the browser tool when you are sizing a cluster, reviewing a performance claim that speedup will track core count, or walking a class through why ninety-five percent parallel still tops out at twenty times. Sweep N upward until speedup approaches the serial limit, then stop buying cores you cannot harvest.

Contrast the result with Gustafson's law if your workload grows with the machine; Amdahl holds the problem size fixed, which is the conservative bound for batch jobs, compilers, and scientific kernels that do not magically enlarge when you add sockets. Results are fully deterministic: the same P and N always yield the same speedup vector, so you can pin examples in tests and lesson materials. There is no language model, no external profiling service, and no randomness.