Cheapest compute
01
—
—
—Empirical single-GPU training planner Snapshot 01
Rent the result,
not the spec sheet.
Compare measured OLM training throughput against current RunPod rates. Find the cheapest route, the fastest route, and every non-dominated option between them.
Full train step
BF16 compute
FP32 state
No checkpointing
2 seeds where complete
01 / Configure the workload
Cost frontier
Cost and time use measured steady-state tokens/second at the selected batch-size objective.
Training tokens
Price assumptions RunPod Secure Cloud
Edit any rate to model a different host or a future RunPod price. Overrides stay in this browser only.
Price snapshot: RunPod Pods,
Fastest finish
02
—
—
—
Pareto choices
03
—
—
Neither slower nor more expensive
Measured fit
04
—
—
—Wall time × compute cost
Training frontier
LOG–LOG
No GPU completed this workload configuration.
02 / Compare every board
Measured options
Median metrics are computed at a shared batch size across the available seeds. Click a row to inspect its raw sweep.
| GPU | Batch | Tokens/s | MFU | Peak VRAM | Replicates | Rate | Time | Compute cost |
|---|
03 / Inspect the evidence
Batch saturation
Every dot below is a measured, stabilized training step—not a theoretical throughput estimate.
Select a GPU
—
No batch sweep is available for this selection.
Per-seed measurements
—| Seed | Batch | Status | Step ms | Tokens/s | Nominal MFU | Configured MFU | Allocated VRAM | Jitter |
|---|
04 / Know what is missing
Coverage matrix
Coverage for the selected model across every tested context. “OOM” means the attempted configuration did not fit.
2-seed complete
1-seed partial
OOM only
Not measured
05 / Read the fine print
Method, not magic
This is an empirical compute-cost comparison, not a promise about end-to-end job billing.
Pick one shared batch
For each GPU, batch metrics are aggregated by median across available seeds. The selected batch maximizes either TPS or configured-clock MFU.
Convert throughput to time
hours = training tokens ÷ measured TPS ÷ 3,600
Apply the rental rate
compute cost = hours × USD per GPU-hour
Keep the boundary honest
Estimates exclude initialization, data loading, storage, checkpoint I/O, evaluation, failures, taxes, and multi-GPU communication. RunPod states that Pods are billed by the minute.