This new technical paper titled “Great Power, Great Responsibility: Recommendations for Reducing Energy for Training Language Models” is from researchers at MIT and Northeastern University.
Abstract: “The energy requirements of current natural language processing models continue to grow at a rapid, unsustainable pace. Recent works highlighting this problem conclude there is an urgent need for methods that reduce the energy needs of NLP and machine learning more broadly. In this article, we investigate techniques that can be used to reduce the energy consumption of common NLP applications. In particular, we focus on techniques to measure energy usage and different hardware and datacenter-oriented settings that can be tuned to reduce energy consumption for training and inference for language models. We characterize the impact of these settings on metrics such as computational performance and energy consumption through experiments conducted on a high performance computing system as well as popular cloud computing platforms. These techniques can lead to significant reduction in energy consumption when training language models or their use for inference. For example, power-capping, which limits the maximum power a GPU can consume, can enable a 15\% decrease in energy usage with marginal increase in overall computation time when training a transformer-based language model.”
Find the technical paper here. Published May 2022.
Related Reading AI Power Consumption Exploding Exponential increase is not sustainable. But where is it all going? 11 Ways To Reduce AI Energy Consumption Pushing AI to the edge requires new architectures, tools, and approaches.
Name* (Note: This name will be displayed publicly)
Email* (This will not be displayed publicly)
Steps are being taken to minimize problems, but they will take years to implement.
But that doesn’t mean it’s going to be mainstream anytime soon.
Companies are speeding ahead to identify the most production-worthy processes for 3D chip stacking.
New capacity planned for 2024, but production will depend on equipment availability.
L5 vehicles need at least 10 more years of development.
Increased transistor density and utilization are creating memory performance issues.
Suppliers are investing new 300mm capacity, but it’s probably not enough. And despite burgeoning 200mm demand, only Okmetic and new players in China are adding capacity.
The industry reached an inflection point where analog is getting a fresh look, but digital will not cede ground readily.
100% inspection, more data, and traceability will reduce assembly defects plaguing automotive customer returns.
Engineers are finding ways to effectively thermally dissipate heat from complex modules.
Different interconnect standards and packaging options being readied for mass chiplet adoption.
Steps are being taken to minimize problems, but they will take years to implement.
AMD CTO Mark Papermaster talks about why heterogeneous architectures will be needed to achieve improvements in PPA.