Gustafson-Barsis Scaled Speedup

The engineering

Gustafson-Barsis answers a different question than Amdahl: instead of fixing the problem size and asking how much faster N processors finish it, it fixes the run time and asks how much bigger a problem those N processors can chew through. In practice engineers scale up the mesh, the sample count, or the resolution when they add nodes — so the serial fraction s here is the serial share of the *scaled* workload, not the original one.

Because the parallel work grows linearly with N, speedup is nearly linear and the pessimism of Amdahl's law disappears. The sanity check: with N=1 speedup is exactly 1, and efficiency tells you how close to ideal you are — if it's dropping fast as N climbs, your serial fraction is the real ceiling, not the core count.

Where this math comes from

In 1988 John Gustafson, then at Sandia National Laboratories, was chasing the Gordon Bell Prize on a 1,024-processor nCUBE hypercube. Amdahl's law predicted that with any meaningful serial fraction such a machine could never deliver more than a modest speedup — yet Gustafson's team was measuring speedups over 1,000. The resolution was that nobody runs a fixed-size problem on a bigger machine; they run a bigger problem in the same time.

Gustafson credited colleague Ed Barsis for the framing and published the reformulation in a one-page Communications of the ACM note, 'Reevaluating Amdahl's Law.' It reframed the whole parallel-computing debate and helped justify the massively parallel machines that followed.

1. 1967Gene AmdahlStates the fixed-size speedup limit that Gustafson later reinterprets.
2. 1988John Gustafson & Ed BarsisPublish scaled speedup after 1,000× results on Sandia's nCUBE.
3. 1988Gordon Bell Prize committeeRecognizes Sandia's massively parallel results, validating the model.

See the full timeline of the math behind every calculator →

Runs entirely in your browser — nothing you enter leaves this page. Your recent runs are stored only on your device.