Accuracy and limits
What to trust, what to treat as an estimate, and what is missing.
NEURAX predicts. Some of what it predicts is checked against PyTorch on published models; some is not yet. Knowing which is which is part of using it well.
Checked against PyTorch#
| Model | Parameters | FLOPs, forward and backward |
|---|---|---|
| GPT-2, Qwen2.5 7B, Mistral 7B | exact | exact |
| BERT, RoBERTa, DeBERTa-v3 | exact | exact |
| T5-small, BART-base | exact | exact |
| Mamba 130M and 2.8B | exact | exact |
| Mixtral 8x7B, Qwen3-30B-A3B, gpt-oss-20b | exact | exact |
Each model is imported from its published configuration, compiled, and compared with what PyTorch itself counts on the same model. These checks run on every build, and a model that drifts fails it. It is not a claim about architectures that are not measured.
Estimated, and now checkable#
- Step time and throughput divide the step's FLOPs by the machine's throughput — measured on this computer, or the published figure of a named card — and its element-wise work by the memory bandwidth. Completed runs at a comparable predicted scale calibrate step time; Training shows each run beside its prediction.
- Kept activations follow each operation's backward rules and are not yet calibrated against measured runs, so peak memory is an estimate.
- Training cost and time follow from those estimates and inherit their uncertainty, plus whatever your cloud actually charges.
- Energy and CO₂ need the machine's power draw and the grid's carbon intensity. Without them they are reported as unknown, never as zero.
Known gaps#
- Diffusion and graph models are drawn rather than imported. Every block of the six families is priced when you draw it; importing a published diffusion pipeline only its UNet is imported, and graph networks are drawn, not imported.
- Multimodal models. Only the language tower is imported, so totals cover the text side alone.
- Not every design trains. A repeated block whose body is also drawn separately cannot be generated faithfully, and NEURAX refuses rather than train a different architecture.
Page 31 of 33. This guide is in English for now. It is the same guide that is built into the NEURAX desktop app (the ? button, or F1). Found a mistake? Tell us.