Lowest compute cost01
—
—
—A practical single-GPU training resource
What should you rent
to train your SLM?
Start with the workload, not peak FLOPs. Get the cheapest and fastest measured option, its best batch size, steady-state tokens/second, nominal MFU, peak VRAM, and expected compute time.
Full train step
BF16 compute
FP32 state
No checkpointing
No torch.compile
2 input replicates
Answer in one glance
The cheapest board changes with model scale.
Best measured max-TPS batch at 2,048-token context, priced using the selected source below.
Building the cross-scale frontier…
“Cheapest” means lowest GPU compute cost per billion training tokens. It excludes storage, data loading, setup time, networking, and multi-GPU scaling.
Interactive planner
Price the workload you actually want.
Performance is measured on RunPod hosts. Switching price source transfers the same measured TPS to another provider’s dated GPU rate; it is a cost projection, not a new performance measurement.
Training tokens
i
Loading provider pricing…
Inspect or edit rates —
Rates are normalized to USD per GPU-hour. Some quotes come from multi-GPU bundles; match quality and product details are shown below. Overrides stay in this browser.
Snapshot: source,
Fastest finish02
—
—
—Cost/time frontier03
—
—
Neither slower nor more expensivePriced measured fits04
—
—
—Wall time × compute cost
Training frontier
LOG–LOG
No provider-priced GPU completed this workload.
Compare every board
Measured options
Median metrics at a batch shared by the available input replicates. Select a row to inspect its raw sweep.
| GPU | Best batch | Tokens/s | Nominal MFU | Peak VRAM | Input replicates | Provider rate | Wall time | Compute cost |
|---|
Inspect the evidence
Batch saturation
Every dot is a measured stabilized full training step, not theoretical throughput.
Select a GPU
—
No batch sweep is available for this selection.
Per-input-replicate measurements
—| Input replicate | Batch | Status | Step ms | Tokens/s | Nominal MFU | Configured MFU | Allocated VRAM | Jitter |
|---|
Know what is measured
Coverage matrix
Coverage for the selected model across every tested context. “OOM” means the attempted configuration did not fit.
2-replicate complete
1-replicate partial
OOM only
Not measured
Bracketed estimate
No measured bracket
Method and boundary
Measured performance. Transferable prices.
This is an empirical GPU compute comparison, not a promise about total job billing or provider-specific end-to-end performance.
Measure full training
BF16-autocast forward, shifted-token loss, backward, and fused AdamW update. Twenty timed steps follow adaptive warmup.
Select a useful batch
Metrics are medians across available input replicates. The selected shared batch maximizes TPS or nominal-reference MFU.
Convert TPS to cost
hours = tokens ÷ TPS ÷ 3,600
cost = hours × USD per GPU-hour
Interpolate only inside the data
TPS(P) = T₀ × (P ÷ P₀)α, between adjacent measured model sizes
Loading held-out error…
Read before deciding
Performance host. All TPS, MFU, and VRAM values were measured on the recorded RunPod benchmark hosts. AWS, Google Cloud, Lambda, and Vast.ai values transfer price only.
Software path. Eager PyTorch/OLM, no torch.compile, no activation checkpointing, and no gradient accumulation. Different kernels or precision policies may move the frontier.
Single GPU. Multi-GPU scaling, availability, startup, networking, storage, data pipeline, and interruption risk are outside this version.