Plan your Compute
with transparent methodology.
A guided estimator grounded in documented formulas, benchmark-informed assumptions, and hardware specifications to translate your research plans into credible GPU-hour requests for clusters like Jean Zay.
This tool is a work in progress and will be continuously calibrated with benchmark evidence from MINERVA and partners contributions .
Input
Questionnaire
Section 1
Workload Intent
Clarifies whether compute is for full-model training, partial adaptation, PEFT, or inference/evaluation.
Captures iteration overhead beyond the single best run.
Section 2
Architecture
Architecture family changes default model sizes and typical data profiles.
Select a representative size band or enter a custom value.
Using 7B parameters for the estimate.
Section 3
Data
Not all datasets have the same per-sample compute cost.
How much data is processed for one epoch (before iteration buffers).
How many times to pass through the entire dataset.
Section 4
Efficiency
The numeric format used for matrix multiplications.
Number of GPUs you intend to use for this run.
Output
Draft Estimate
Axis in GPU-hours (auto-scaled per scenario)
Target Cluster
8× H100 (80gb)
Estimated Real Time
32.1 days
GPU-Hours On Recommended Type
6,163 GPU-hours (H100)
Recommended GPU Type
H100 (80gb)
Call Type
Dynamic access
Dynamic limit for H100: 12,500h
“High compute or communication pressure: H100 is recommended for faster kernels and stronger scaling across nodes.”
Fits with multi-GPU sharding; communication/memory overhead can increase.
Info
This setup spans 2 nodes on NVIDIA H100 SXM5 (4 GPUs/node).
Inter-node communication over IB400 can reduce scaling efficiency.