Engineering | Cost analysis
Before renting GPUs to train PetSpeak's bark interpreter, we ran the numbers. Here is the honest breakdown: what one training run actually costs, and why the biggest GPU is the wrong buy for a model this size. Grounded in a measured throughput benchmark of our real pipeline and live RunPod pricing (July 2026).
The bottom line
$0.33
Per run on a right-sized GPU (A40 / RTX 3090)
~35 min
Wall-clock per run, billed per-second
~$7
To run 20 tuning experiments
One end-to-end run downloads ~15,000 real dog barks, auto-labels them with an AudioSet-pretrained teacher model, trains our classifier, and exports the deployable model for web and iOS. Here is where that money goes.
Cost per run
RunPod on-demand rates observed July 2026, priced against a conservative 0.75-hour bill. The bars are relative cost — the analysis is visible at a glance.
Best value mid-tier 24–48 GB cards. Overkill flagship GPUs cost 6–18× more for zero benefit on a model this small — a $0.44/hr A40 finishes in the same wall-clock as a $2.99/hr H100, or a $5.89/hr B200.
Where the time goes
Measured: our AudioSet teacher labels barks at 0.80 clips/sec on a laptop CPU — about 5.2 hours for 15,000 clips. That five-hour gap is the entire case for renting a GPU: the labeling model is a transformer, and transformer batch inference is exactly what GPUs accelerate, dropping labeling to a few minutes.
The takeaway
The instinct is to reach for the biggest GPU. For a model this size, the data says the opposite: a $0.44/hr card matches the flagship on wall-clock and wins decisively on cost per run. Right-sizing — not raw horsepower — is the analysis. Cheap runs are why we iterate on the model instead of rationing GPU time.
Method: CPU throughput measured in-repo; GPU runtimes estimated from that baseline plus standard transformer-inference speedups, against a conservative 0.75-hour bill. Prices are RunPod on-demand rates observed July 2026 (runpod.io/pricing); Community Cloud and spot instances run cheaper still. A ~5 GB storage volume is ≈ $0.25/month, negligible per run. Re-check current rates before quoting.