Skip to main content
Category: Laws
Type: Parallel computing and systems scaling law
Origin: Gene Amdahl, AFIPS Spring Joint Computer Conference (1967)
Also known as: Amdahl bound; fixed-size speedup limit; strong-scaling ceiling
Quick Answer — Amdahl’s Law says the speedup you can get from parallel resources is capped by the part of the work that must stay serial. If a fixed job is 5% sequential, even infinite processors cannot beat a 20× overall speedup. The practical move is to shrink serial bottlenecks before buying more parallelism.

​
What is Amdahl’s Law?

Amdahl’s Law is the rule that overall system speedup is limited by the fraction of work that cannot be parallelized, when problem size is held fixed.
The effort expended on achieving high parallel processing rates is wasted unless it is accompanied by achievements in sequential processing rates of very nearly the same magnitude.
Gene Amdahl’s 1967 argument was blunt: for real workloads with irregularities and coordination overhead, the serial remainder eventually dominates. In the modern formula, if fraction p of a fixed job can run in parallel on N processors (and 1 − p stays serial), speedup approaches 1 / (1 − p) as N grows—not N.

​
Amdahl’s Law in 3 Depths

  • Beginner: Speeding up only part of a task cannot make the whole task infinitely fast.
  • Practitioner: Measure the serial fraction first; optimize that path before adding cores, workers, or servers.
  • Advanced: Treat Amdahl as a strong-scaling bound for fixed work; switch to scaled-problem thinking when users grow the job with capacity (Gustafson-style weak scaling).

​
Origin

Computer architect Gene Amdahl, then at IBM, presented “Validity of the single processor approach to achieving large scale computing capabilities” at the 1967 AFIPS Spring Joint Computer Conference. The short paper had no famous equation, but argued that multiple-processor machines would struggle to deliver peak performance on irregular real problems because sequential work and coordination overhead remain. Later teaching distilled the idea into the familiar bound: for fixed problem size, speedup S(N) = 1 / ((1 − p) + p/N). Classic classroom numbers make the ceiling vivid—if p = 0.95, the infinite-N limit is 20×; with 64 processors the same p yields roughly 15×, not 64×. In 1988, John Gustafson (with work at Sandia National Laboratories) reframed the debate in “Reevaluating Amdahl’s Law”: scientists often scale problem size with machine size and keep runtime roughly constant. On a 1024-processor system, Sandia reported scaled speedups on the order of ~1020–1022 for three applications—near-linear when the parallel work grows with N. That does not refute Amdahl for fixed jobs; it changes the question from “same work, less time” to “more work, same time.”

​
Key Points

Amdahl’s Law is a budgeting tool for where speedup can and cannot come from.
1

Serial work sets the ceiling

Whatever must run one-at-a-time—initialization, locking, final aggregation, human approval—bounds total improvement. Infinite parallel capacity cannot erase that fraction.
2

Extra processors show diminishing returns

Each added core helps only the parallel slice. As N rises, returns shrink toward 1/(1 − p), a form of diminishing returns under fixed workload.
3

Fix the bottleneck before you scale out

Buying more machines while a single-threaded path or shared lock owns 10–20% of runtime wastes budget. Profile first; parallelize second.
4

Know which scaling question you are asking

Fixed-size speedup (strong scaling) is Amdahl territory. Growing the problem with capacity (weak scaling) needs a different scorecard—often closer to Gustafson’s scaled-speedup view.

​
Applications

Use the law whenever someone proposes “just add more parallelism” without naming the serial fraction.

Software performance

Before multi-threading a hot path, estimate what share of wall time is still single-threaded I/O, GC pauses, or lock contention.

Cloud and capacity planning

Model request pipelines: if auth or a shared database write is serial, horizontal pods will plateau until that stage is redesigned.

Team and process design

Treat one approver or one integration gate as a serial fraction; adding people elsewhere will not shrink end-to-end cycle time past that limit (Brooks’s Law is a cousin on coordination cost).

Learning and personal workflow

Parallelize research reading, but keep a serial “decide and write” block; more tabs will not beat a blocked decision step.

​
Case Study

Amdahl’s fixed-size math and Gustafson’s Sandia results together show when the law bites. Suppose 95% of a batch job is parallelizable. Amdahl’s Law says even infinite processors cannot exceed a 20× speedup on that same job; with 64 processors the predicted speedup is about 15.4×. That is the strong-scaling story: same work, hope for less time, serial remainder wins. In 1988, Gustafson and colleagues at Sandia reported results on a 1024-processor hypercube for three scientific applications (including beam stress analysis and fluid dynamics workloads). They measured scaled speedups of roughly 1021, 1020, and 1022—near the processor count—because researchers enlarged the problem (for example, finer grids) so parallel work grew with the machine while serial overhead stayed comparatively small. The lesson is precise: Amdahl’s Law correctly caps fixed-size speedup; it is misused when you treat it as a ban on large parallel machines for scaled science. Choose the metric that matches the goal—finish the same job faster, or finish a bigger job in similar time.

​
Boundaries and Failure Modes

Amdahl’s Law assumes a fixed problem and an idealized split into serial and perfectly parallel parts. Real systems add communication, load imbalance, and memory bandwidth limits that can make achieved speedup worse than the formula. It fails as a dismissal tool when the workload can grow usefully with capacity—image resolution, simulation fidelity, or batch volume. In those cases, weak scaling may still deliver large value even if fixed-size speedup looks modest. A common misuse is quoting the infinite-N ceiling to reject any parallel investment without measuring p. If the serial fraction is 0.5%, the ceiling is 200×; the law then argues for parallelism, not against it.

​
Common Misconceptions

Clear use requires separating the formula, the scaling goal, and neighboring laws.
No. It means returns are bounded by the serial fraction for fixed work. High p still justifies substantial parallel investment.
No. They answer different questions: fixed problem size versus scaled problem size. Both can be true in their domains.
No. Any system with a non-parallelizable stage—network fan-in, single-writer stores, human review—obeys the same ceiling logic.
These pages help place Amdahl’s Law among hardware trends, team scaling, and return curves.

Moore's Law

Hardware density trends raise capacity; they do not remove serial bottlenecks by themselves.

Brooks's Law

Adding people can increase coordination cost—another serial-like drag on delivery.

Diminishing Returns

Extra inputs yield smaller gains once a constraining factor dominates.

Wirth's Law

Software complexity can spend hardware gains faster than chips deliver them.

Hofstadter's Law

Complex work takes longer than expected, even when you expect that.

Metcalfe's Law

Network value scales with connections—useful contrast when growth, not serial fraction, is the story.

​
One-Line Takeaway

Before you buy more parallelism, measure the serial fraction—that number is your real speedup ceiling for fixed work.