Fire is a GPU kit that exists outside of computers.
Fire is NOT energy efficient and only runs ONCE.

In a situation where mineral and chemical based GPUs run out, Fire as a wood based GPU is the only type available for AI training.

USAGE:

One Fire GPU kit contains 4 panels of wooden equilateral triangles.


1) Etch on three panels your training script, neural network model, and dataset.

2) Assemble the triangles together into a tetrahedron. Align the edges based on the number specified. Edges with the same numbers should be aligned together.

3) Use threads to tie the vertices and hold the tetrahedron together.

4) Put combustion materials inside the tetrahedron through holes on one of the wood panels.

5) Combust the tetrahedron. Make sure enough oxygen is in contact with the tetrahedron in the process of combustion.

PHOTO DOCUMENTATION OF OBJECTS:

CAUTION:

a) Unlike a mineral and chemical based GPU that can run a million iterations over thousands of data inputs for AI to learn, ONE Fire GPU kit can only run ONE iteration of training with a limited number of data specified.

b) Data to be trained with should be carefully selected as Fire holds only a small number space for data inputs. Fire can hold 196, 324, or 784 pieces of data, and the size of the capacity corresponds to the size of the wood to be burned.

c) AI training being performed with Fire is unpredictable due to possible deficiencies of your training codes and certain unpredictability of neural network models.

d) The use of Fire is limited since running Fire is detrimental to your local environment. Please double check your programs and datasets to prevent errors.

e) Avoid using Fire for only test purposes. Decisions to run Fire should be carefully made.

f) Check the local air conditions report before running Fire and make sure you don’t go over the daily carbon dioxide emission limits.

SAMPLES (scans of etchings):

FIRE196 the training and model scripts are based on GPT2. It accommodates memory IDs of 196 sample text data.

FIRE324 the training and model scripts are based on DCGAN. It accommodates memory IDs of 324 sample image data.

FIRE784 the training and model scripts are based on Open Jukebox. It accommodates the memory IDs of 784 sample sound data.

INTERACTIVE DEMO:

[ Download Demo Windows / Mac ]

The demo briefly shows the procedures of using Fire. It is by no means a substitute to the functionality of the actual wood based GPU.

A video of the interactive demo is also provided below.