Skip to content
GIRIVON
Infrastructure+
Systems+
Components+
Computing Devices+
Solutions+
Compute Tools+
Knowledge+

HPC

Amdahl's law and HPC cluster sizing

Why adding cores stops helping, how Gustafson's law changes the picture, and how to size a cluster from demand.

GIRIVON Engineering · Updated 19 Nov 2026 · 1 min read

If a fraction p of a program's work parallelises perfectly, Amdahl's law gives the speedup on n processors:

S(n) = 1 ÷ ((1 − p) + p ÷ n)

The serial part sets a hard ceiling of 1 ÷ (1 − p). With p = 95%:

  • on 64 processors: 1 ÷ (0.05 + 0.95 ÷ 64) ≈ 15.4× — a parallel efficiency of about 24%
  • on infinitely many processors: at most 20×

Gustafson's view

In practice, larger machines are usually used to solve larger problems in the same time. Gustafson's law gives the scaled speedup S(n) = n − (1 − p)(n − 1): about 60.9× on 64 processors for p = 95%. Both views are correct; they answer different questions.

Measuring p

Time the same job at two processor counts, or profile it. Real scaling is also limited by communication, load imbalance and memory bandwidth, so measured efficiency is usually below Amdahl's prediction.

Sizing a cluster

  1. Estimate demand in core-hours or node-hours per month from your applications.
  2. Choose node design for the bottleneck — many CFD and finite-element codes are memory-bandwidth-bound, so populate every memory channel.
  3. Compute nodes from cores, memory and GPUs at once; the binding constraint decides.
  4. Check theoretical peak (cores × clock × FLOPs per cycle), then racks, power and cooling.

Work through these steps with the Parallel Computing Estimator, the Compute Node Calculator and the HPC Cluster Planner.