What NEURAX does for you

The questions it answers, and the ones it does not.

NEURAX answers questions about a neural network before you train it. You describe an architecture — by drawing it, loading a template, or importing a config — and it tells you what building that model would cost: how much VRAM it needs, how many FLOPs it takes, how long training would run, what that costs in dollars and in carbon.

Answers arrive in well under a second, with no GPU, and the same design always gives the same numbers.

What it will not do#

• It will not answer by running your model. Every figure in Simulation is computed from the graph. Nothing is sampled, nothing is executed, and the same design always gives the same numbers.
• It will not guess at your hardware. If it cannot read a specification for the card it found, it measures the machine instead, and says which figures are approximate.
• It will not train a design it cannot generate faithfully. Where the generated model would not match the one that was analysed, it refuses and names why, rather than training something else.

What it does beyond the analysis#

• It reads the machine it runs on. Processor, memory, disk and any accelerator, and it measures what that hardware actually sustains.
• It generates the design as PyTorch. A real nn.Module, checked against the analysis before anything is run.
• It trains, and checks itself. A run on your own machine reports its real parameter count, memory and speed beside what was predicted.

Page 1 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.