A performance law that relates the maximum possible overall speedup of a system to the fraction of the workload that can be parallelized and the speedup of that parallel portion; it shows diminishing returns when improving only a subpart of computation under a fixed problem size.

Principle

Total speedup is limited by the serial (non-parallelizable) fraction: S_total ≤ 1 / ((1 - p) + p / S_p), where p is the parallelizable fraction and S_p is the speedup of the parallel part. As the number of processors grows, the serial fraction dominates the limit.

Demonstration

If 90% of a program is parallelizable (p = 0.9), then even with infinite processors the maximum speedup is 1/(1-0.9) = 10; making the parallel part infinitely fast still yields at most a 10× overall improvement.

Misapplication

Applying Amdahl's law to scenarios that scale the problem size with more processors (weak scaling) or ignoring communication and synchronization overheads leads to incorrect conclusions about achievable speedup.

Consequence

Directs optimization efforts toward reducing the serial fraction and identifying true bottlenecks; predicts diminishing returns for adding parallel resources without reducing sequential work.

Reversal

Gustafson's law provides a complementary perspective by holding execution time constant and increasing problem size, showing that perceived scalability can grow with workload size—contrasting Amdahl's fixed-size assumption.

Boundary

Assumes a fixed total problem size and that the proportions of serial and parallel work are known and constant; excludes overheads from communication, load imbalance, memory contention, and changes in algorithmic complexity with scale.

Semantic Tension

Often contrasted with Gustafson's law; Amdahl emphasizes limits under strong scaling (fixed problem), which can be misread as a universal cap on parallelism when the workload or conditions change.

Synthesis

Amdahl's law formalizes the limit imposed by sequential portions of work on parallel speedup for a fixed-size problem: it warns that improving only the parallel part or increasing processors yields diminishing returns unless the serial fraction is also reduced.