模拟内存计算注意力机制,用于快速和节能的大型语言模型
Nathan Leroux1, Paul-Philipp Manea2,3, Chirag Sudarshan4
1PGI-15, Forschungszentrum Jülich, Jülich, Germany. n.leroux@fz-juelich.de.
Nature computational science
|September 8, 2025
概括
研究人员开发了一个用于生成变压器的新型内存计算架构. 这种设计通过使用增益单元进行自我注意计算,在大型语言模型中显著降低了延迟和能耗.
科学领域:
- 计算机科学 计算机科学
- 电气工程 电气工程
- 人工智能的人工智能
背景情况:
- 变压器网络和自我注意机制是大型语言模型 (LLM) 的基础.
- 当前的发电变压器面临着由于图形处理单元 (GPU) 和静态随机访问存储器 (SRAM) 之间的数据移动而导致的延迟和能源瓶.
研究的目的:
- 引入定制的自我注意内存计算架构,以克服生成变压器中现有的延迟和能量限制.
- 为了实现高效,低功耗的计算,LLMs.
主要方法:
- 开发了一种新的内存计算架构,利用基于电荷的内存增益单元进行并行模拟点产品计算.
- 设计了一个初始化算法,以解决模拟增益电池电路中的非理想性问题,使得性能与未经完全重新训练的GPT-2等预训练模型相提并论.
主要成果:
- 与基于GPU的系统相比,拟议的架构将注意力延迟降低高达2个数量级,能源消耗降低高达4个数量级.
- 使用新的架构和初始化算法,实现了与GPT-2可比的文本处理性能.
结论:
- 定制内存计算架构为超快速和低功耗的发电变压器提供了实质性的进步.
- 这种方法为更高效和更容易获得的大型语言模型部署铺平了道路.
相关概念视频
Understanding Memory
1.3K
Memory is the retention of information or experiences over time, facilitated through three main processes: encoding, storage, and retrieval. Encoding is the process of inputting information into the memory system. For instance, when listening to a lecture, watching a play, reading a book, or having a conversation, the brain is actively encoding information. This initial stage involves transforming sensory input into a form that can be processed and stored by the brain. Various factors, such as...
1.3K
Mechanistic Models: Compartment Models in Algorithms for Numerical Problem Solving
292
Mechanistic models play a crucial role in algorithms for numerical problem-solving, particularly in nonlinear mixed effects modeling (NMEM). These models aim to minimize specific objective functions by evaluating various parameter estimates, leading to the development of systematic algorithms. In some cases, linearization techniques approximate the model using linear equations.
In individual population analyses, different algorithms are employed, such as Cauchy's method, which uses a...
In individual population analyses, different algorithms are employed, such as Cauchy's method, which uses a...
292
Long-Term Memory
654
Long-term memory is a relatively permanent type of memory, capable of storing vast amounts of information over extended periods. Its storage capacity is generally considered unlimited.
Long-term memory can be categorized into two primary types: explicit and implicit memory. Explicit memory, also known as declarative memory, involves the conscious recollection of information that we deliberately try to remember, recall, and articulate. This type of memory encompasses specific facts, events, and...
Long-term memory can be categorized into two primary types: explicit and implicit memory. Explicit memory, also known as declarative memory, involves the conscious recollection of information that we deliberately try to remember, recall, and articulate. This type of memory encompasses specific facts, events, and...
654
System of Memory
7.2K
Memory is categorized into three major systems: sensory memory, short-term memory (STM), and long-term memory (LTM). These systems differ in their capacity and the duration for which they can hold information. Sensory memory captures raw sensory input from the environment, holding it for just a few seconds or less. For example, on hearing a brief, loud sound, like a car horn honking, the sound seems to linger in the mind for a moment even after it stops. This is an instance of sensory memory...
7.2K
Language and Cognition
716
Language serves as a bridge between ideas and communication, influencing how individuals perceive and interact with the world. Psychologists have long debated whether language shapes thought or vice versa. This discussion gained grip with Edward Sapir and Benjamin Lee Whorf in the 1940s, who proposed that language determines thought, a concept known as linguistic determinism. They suggested that the vocabulary and structure of a language influence how its speakers think and perceive reality.
716
Improving Translational Accuracy
14.1K
Base complementarity between the three base pairs of mRNA codon and the tRNA anticodon is not a failsafe mechanism. Inaccuracies can range from a single mismatch to no correct base pairing at all. The free energy difference between the correct and nearly correct base pairs can be as small as 3 kcal/ mol. With complementarity being the only proofreading step, the estimated error frequency would be one wrong amino acid in every 100 amino acids incorporated. However, error frequencies observed in...
14.1K

