3× A40 vs 1× RTX 5090
DeepLab, 63M parameters · 300 epochs in total · identical final result on both setups.
3.04×faster221.4 h → 72.8 h of wall-clock time
−47%cost of the run$219.17 → $115.29, our fee included
$103.88saved148.6 hours, six days back
3× A40 on ResonTech1× RTX 5090
Wall-clock time · shorter is better
3× A40
72.8 h
1× 5090
221.4 h
Cost of the run · lower is better
3× A40
$115.29
1× 5090
$219.17
The A40 run includes the 20% ResonTech fee; the 5090 is its raw hourly rate.
Hourly rate
3× A40
$1.32
1× 5090
$0.99
3 × $0.44 for the A40s: 33% more per hour than the 5090.
Hourly rate is the wrong metric. Three A40s cost 33% more per hour than one RTX 5090, and finish the same job 3× sooner, so the total bill is 47% lower.
Six days back. The 5090 run took 9 days 5 hours. The A40 cluster finished in about 3 days: the same 300 epochs and the same result.
In numbers
How we measured
- Wall-clock and cost come from the jobs’ own timestamps and the on-demand prices at the time of the run (the A40 run: 29 May → 1 June 2026).
- Same model, same data, the same 300 epochs; both setups reached the same final result.
- The A40 run is federated across three workers on ResonTech. Its cost includes our 20% fee; the RTX 5090 is shown at its raw hourly rate, which makes the −47% the conservative figure.
What this does not show
- Per-GPU speed or scaling efficiency. The A40 and the RTX 5090 are different cards. This compares two setups for the same job.
- Linear scaling. Three GPUs finishing 3.04× sooner is not a claim that ten GPUs finish ten times sooner.
Federated accuracy is covered in Accuracy is preserved, and the tuning behind it in How FL works on ResonTech.