Skip to content

jelspace/PRECOMPUTED-AI-WEIGHTS

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 

Repository files navigation

PRECOMPUTED-AI-WEIGHTS

PRECOMPUTED AI WEIGHTS

MAY WORK SOMEHOW FOR EXAMPLE IF IN/OUT NEURON DATA IS 1b-16b(NOT TESTED)

SCALABLE MAY WORK SERIALIZED AND IN PARALLEL IN ALL MICROCONTROLLERS / FPGAS ,SPECIALIZED ASICS


!!!!!! INNOVATION IS FABULOUS BECAUSE IS SCALABLE MAY WORK SERIALIZED AND IN PARALLEL IN MANY DEVICES !!!!!!


BY (1D,2D,3D,..?)CHUNKS THEN KAN OR DIRECT(VERY SMALL MODELS) , ALL Q4 POSSIBILITIES/COMBINATIONS ARE 100TB FOR EXAMPLE FOR 10B 20B MODEL (DATABASE SIZE MAY WRONG COMPUTED 100TB WITH PRECOMPUTED WIEGHTS)(THIS WAS THE FIRST AND DREAMING IDEA WITH ALL WEIGHTS PRECOMPUTING)

NOT TESTED JUST IDEA


BENEFITS: VERY LOW POWER CONSUMPTION (COMPARED WITH OLD COMPUTE ALL WEIGHTS TECHNOLOGY,MULTIPLICATION NEEDS MORE POWER THAN ADDING PROBABLY NOT TESTED)

SUPPORTS ALL TEXT,VISION,STT,TTS,DIFFUSION VIDEO/AUDIO/3D GANERATION,SIMULATORS MODELS


====================== THIS IS MAIN IDEA ======================

MAIN IDEA IS TO PRECOMPUTE ONLY MULTIPLICATION(DIV) AND STORE RESULTS IN DATABASE(NEW MODEL FILE),FOR ALL POSSIBLE INPUTS OF EACH NEURON (4b 16 COMBINATIONS,8b 256 COMBINATIONS,....) NEW MODEL FILE SIZE WILL BE OLD SIZE(Q BASED) X RESOLUTION NEURON DATA IN


!!!!!!!!!! USE MANY IC DEVICES TO COMPUTE !!!!!!!!!!


INSTANT RESULT!!!!!!!!!!

AI ASIC FLASH IS POSSIBLE BY COMBINING INTERNALY FLASH IC WITH ADD LOGIC THAT WILL COMPUTE FOR 1 CLOCK OR FEW CLOCKS ENTIRE MLP NEURAL NET , MEAN ALMOST FLASH COMPUTING OF THE TOKEN FOR ONE CLOCK THIS WAY 200 MILLION TOKENS PER SECOND ARE POSSIBLE AT 200MHZ CLOCK


SITUATION IS VERY INTERESTING IN Q2(Q3?/Q4?) BECAUSE POSSIBILITY TO PRECOMPUTE ALSO AND ADDING ,IN ADDING LEVELS BY 2 ARE FORMED,I THINK COMPUTATIONS ARE 2 TIMES LOWER SO SPEED IS TWICE FASTER AT EACH LEVEL BUT SIZE IS UP ^X NOT SURE IN THIS COMPUTE IT JUST BY MIND.....


!!!!!! INNOVATION IS FABULOUS BECAUSE IS SCALABLE MAY WORK SERIALIZED AND IN PARALLEL IN MANY DEVICES !!!!!!


OTHER MAIN ADVANTAGE IS POSSIBILITY TO FINE TUNE THE PRECOMPUTED DATABASE WHICH IS ULTRA FINE TUNING OF THE NEURAL NET NEVER POSSIBLE WITH MULTIPLICATION OF ALL IN NEURON DATA COMBINATIONS (FOR ONE NEURON OUTPUT) WITH ONE FIXED WEIGHT VALUE ONLY NEURAL NETS ! WILL UP THE QUALITY OF SMALL NETS TO THE SKY ( NOT TESTED )


IDEA IS NOT TESTED AND NOT FULLY FINISHED

About

PRECOMPUTED AI WEIGHTS,INSTANT RESULT

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published