A crew of engineers at AI inference expertise firm BitEnergy AI experiences a way to scale back the power wants of AI purposes by 95%. The group has printed a paper describing their new method on the arXiv preprint server.
As AI purposes have gone mainstream, their use has risen dramatically, resulting in a notable rise in power wants and prices. LLMs similar to ChatGPT require loads of computing energy, which in flip means loads of electrical energy is required to run them.
As only one instance, ChatGPT now requires roughly 564 MWh each day, or sufficient to energy 18,000 American properties. Because the science continues to advance and such apps develop into extra widespread, critics have instructed that AI purposes could be utilizing round 100 TWh yearly in just some years, on par with Bitcoin mining operations.
On this new effort, the crew at BitEnergy AI claims that they’ve discovered a option to dramatically scale back the quantity of computing required to run AI apps that doesn’t end in diminished efficiency.
The brand new method is fundamental—as an alternative of utilizing advanced floating-point multiplication (FPM), the tactic makes use of integer addition. Apps use FPM to deal with extraordinarily massive or small numbers, permitting purposes to hold out calculations utilizing them with excessive precision. It is usually probably the most energy-intensive a part of AI quantity crunching.
The researchers name their new technique Linear-Complexity Multiplication—it really works by approximating FPMs utilizing integer addition. They declare that testing, to this point, has proven that the brand new strategy reduces electrical energy demand by 95%.
The one downside it has is that it requires completely different {hardware} than that at the moment in use. However the analysis crew additionally notes that the brand new kind of {hardware} has already been designed, constructed and examined.
How such {hardware} can be licensed, nonetheless, continues to be unclear—at the moment, GPU maker Nvidia dominates the AI {hardware} market. How they reply to this new expertise may have a serious influence on the tempo at which it’s adopted—if the corporate’s claims are verified.
Extra info:
Hongyin Luo et al, Addition is All You Want for Power-efficient Language Fashions, arXiv (2024). DOI: 10.48550/arxiv.2410.00907
© 2024 Science X Community
Quotation:
Integer addition algorithm may scale back power wants of AI by 95% (2024, October 12)
retrieved 14 October 2024
from https://techxplore.com/information/2024-10-integer-addition-algorithm-energy-ai.html
This doc is topic to copyright. Aside from any honest dealing for the aim of personal examine or analysis, no
half could also be reproduced with out the written permission. The content material is offered for info functions solely.
