Nemotron Post-training Calculator

How to use this calculator

Choose the model and GPU in the Model & GPU section. The calculator will populate a starting scenario; review the workload assumptions in the next step before using the estimate.
Refine the customization workload and, if needed, Advanced Configurations. Keep the defaults for an early estimate, or enter the workload and technical settings you expect to use.
Read the results. Review the GPUs needed and estimated wall time. Enter your own hourly GPU rate in the cost card if you also need a compute-cost estimate.

Important: This calculator is intended to provide a directional planning estimate—not a performance guarantee, procurement quote, or service-level commitment.

Inputs

Model & GPU

Start by choosing the model to customize and the GPU you expect to rent or own.

Model

GPU

Customization Workload

Use the defaults for an early estimate, or replace them with your tokenized dataset and planned epoch count.

Training approach

Maximum sequence length

Training dataset size (million tokens)

Epochs

Advanced ConfigurationsOptional controls for refining the estimate; leave these at defaults for directional planning.

Effective attended length

Benchmark planning factor (%)

MFU override (%)

Runtime buffer (%)

Cost buffer (%)

GPUs needed

—

Estimated wall time

—

GPU compute cost estimate

—

GPU rental price (USD/GPU-hour)

Verified Training Setting

Policy-training setting

Tensor (TP)—

Pipeline (PP)—

Context (CP)—

Expert (EP)—

Verified memory-fit starting point. See Methodology and assumptions for limitations.

Calculation trailShow formulas, definitions, units, and where the values come from.

Unit check: achieved compute rate TFLOP/s/GPU ÷ model FLOPs per token GFLOP/token = token throughput tokens/s/GPU.

A FLOP is one floating-point operation. FLOP/s is a compute rate. 1 TFLOP = 1,000 GFLOP. “/GPU” remains in the rate until per-GPU throughput is multiplied by the active GPU count.

Measured Benchmark CalibrationShow the selected measured run used to calibrate the speed assumption.

Benchmark	Reported tokens/s/GPU	Reported MFU	Source-formula FLOPs check	Reproduced tokens/s/GPU

MFU means Model FLOPs Utilization. Reference MFU is recomputed from measured tokens/s/GPU, model FLOPs/token, and the GPU’s dense BF16 peak.

Speed-assumption sensitivity

Model FLOPs Utilization	Wall time	Cost

Higher Model FLOPs Utilization (MFU) means the GPU completes more model math each second, reducing time and compute cost.

LoRA vs full SFT defaults

Approach	GPUs	Time	Cost

Comparison uses each method’s own dataset-token and epoch defaults plus its verified topology. Full SFT can finish sooner only when its memory-fit topology uses many more GPUs; compare GPU-hours and compute cost as well as wall time.