Does doubling GPUs halve the time?

No. Communication overhead means the speedup is sublinear, so the cost per useful GPU hour rises with cluster size even though wall-clock time falls.

Should I budget for one run?

No. The first attempt rarely produces a usable model, and budgeting for a single run is the most common way training costs overrun.

Technologyai

GPU Training Cost and Time Calculator

GPU training cost with scaling efficiency and failed runs counted. Doubling GPUs does not halve training time, so the cost per useful hour rises with cluster size.

Also called: training compute cost, gpu hour calculator.

Total training cost

₹4,75,200

₹4,75,200 for 3 runs of 79.2 hours on 8 GPUs. A single run costs ₹1,58,400. Scaling at 85% means 285.1 GPU hours are lost to coordination, worth ₹71,280. Scaling efficiency below ninety percent means a meaningful share of the spend buys coordination rather than computation.

Cost per run

₹1,58,400

Hours including overhead

79.2

Total GPU hours

1,900.8

GPU hours lost to scaling

285.1

Cost of imperfect scaling

₹71,280

Saving from spot pricing

₹0

Effective cost per useful GPU hour

₹294

On multi-GPU scaling

Scaling efficiency below ninety percent means a meaningful share of the spend buys coordination rather than computation.

On prices

Prices are inputs, not constants. Model pricing changes often, so take the current rate from the provider rather than relying on a default here.

How this is calculated

Method and background

How this is calculated

The bill is GPU count times hours times rate, multiplied by the number of runs you actually need. Two realities inflate it. Multi-GPU scaling is never perfect: communication overhead means eight GPUs deliver perhaps six and a half GPUs of useful work, and the gap is paid for. And the first run rarely works, so budgeting for one run understates by the failure rate. Spot pricing cuts the rate substantially and adds restart risk, which is why checkpointing overhead is a separate input.

imperfect scaling means the wall-clock saving from more GPUs is less than proportional, and the cost is not

n: GPU count
eta: Scaling efficiency

Worked examples

Each of these is asserted on every build. If a change to the engine ever moved one of these answers, the build would fail before the page could print it.

three runs on eight GPUs

GPUs: 8
Cost per GPU hour: ₹250
Training hours: 72
Multi-GPU scaling efficiency: 85%
Runs expected including failures: 3
Spot discount if used: 0%
Time lost to checkpointing and restarts: 10%

Total training cost₹4,75,200

72 x 1.1; 8 x 79.2 x 250

Open this example

spot pricing halves the rate

GPUs: 8
Cost per GPU hour: ₹250
Training hours: 72
Multi-GPU scaling efficiency: 85%
Runs expected including failures: 3
Spot discount if used: 50%
Time lost to checkpointing and restarts: 10%

Total training cost₹2,37,600

boundary

Open this example

Method and limits

What it assumes

A constant scaling efficiency, which typically falls as the cluster grows.

What it deliberately does not model

Scaling efficiency depends heavily on model architecture, batch size and interconnect.
Spot preemption can lose work between checkpoints entirely.
Data preparation and evaluation compute are not included.

Formula version 1.0.0 · definition 1.0.0 · India · Report a problem with this calculator

Frequently asked questions

Does doubling GPUs halve the time?: No. Communication overhead means the speedup is sublinear, so the cost per useful GPU hour rises with cluster size even though wall-clock time falls.
Should I budget for one run?: No. The first attempt rarely produces a usable model, and budgeting for a single run is the most common way training costs overrun.