Amdahl Law

Amdahl’s Law tells you the best possible overall speedup when only part of a workload can be accelerated. It works by splitting the original runtime into the fraction that stays serial, 1-p, and the fraction p that becomes s times faster, so the new total time is the sum of those two contributions: (1-p) + p/s. Taking the reciprocal gives the total speedup relative to the original single-core or unoptimized runtime.

Use this when analyzing the benefit of parallel processors, GPUs, SIMD, or any optimization that speeds up only part of a program. It is especially useful for estimating diminishing returns and for deciding whether improving the parallel portion is worth the engineering effort.

• Treating the speedup s as the overall speedup instead of the speedup of only the parallel fraction
• Using p as a percentage without converting it to a decimal, such as writing p = 80 instead of p = 0.8
• Assuming the total speedup can grow without limit even when a nonzero serial fraction remains