Neuralwatt首席执行官发布研究报告，深入分析AI代理框架对语言模型能源效率的影响1。研究基于2026年9月8日至10月13日期间对OpenCode、Pi、omp、Kilo Code等开源代理的测试1，发现不同框架在提示词缓存性能上存在显著差异。
缓存命中率的细微差别造成的能耗差异令人瞩目1。当缓存命中率为92%时，每100个token中8个未缓存；而97%的缓存命中率下仅有3个未缓存token，两者间能耗差异达2.7倍1。在普通请求中，缓存未命中会使单次请求能耗增加约2倍，在GLM-5.3模型上甚至增加约2.5倍1。对于超过128k tokens的长对话，单次重读超过50%提示词的请求消耗相当于约600次正常请求的计算量1。
研究指出，代理框架在构建提示词时改动已发送内容是导致缓存破坏的主要原因1。具体包括OpenCode的计划模式提醒、文件修剪、推理难度切换、MCP服务器连接断开、技能添加、日期变更及提示词压缩等因素1。其中，工具列表或努力级别变更导致缓存破坏概率提升20倍，约占所有重读token的19%1。通过优化代理设计可显著改善能源浪费问题。
Neuralwatt CEO Chad Gibson released research analyzing how AI agent frameworks affect the energy consumption of language models through prompt caching behavior 1. The study, conducted between September 8 and October 13, 2026, examined open-source agents including OpenCode, Pi, omp, and Kilo Code, revealing significant efficiency disparities across different architectures 1.
Cache hit rates among tested frameworks ranged from 92% to 97%, corresponding to a 2.7-fold difference in energy consumption per token 1. When cache misses occur, individual requests consume approximately 2 times more energy in typical scenarios, and up to 2.5 times more energy when processed through the GLM-5.3 model 1. In extended conversations exceeding 128,000 tokens, single requests requiring re-reading over 50% of the prompt context consume computational resources equivalent to approximately 600 normal requests 1.
The research identified that most cache destruction stems from how agent frameworks modify previously sent content during prompt construction 1. Key causes include planning mode notifications in OpenCode, file pruning in omp, reasoning difficulty switches, MCP server connections and disconnections, skill additions, date changes, and prompt compression 1. Changes to tool lists or effort levels increase the probability of cache destruction by roughly 20 times, accounting for approximately 19% of all re-read tokens in production environments 1. The findings suggest that optimizing agent design can significantly mitigate these energy inefficiencies 1.
评论
还没有评论，欢迎留下第一条。