[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-62efd1eb-8dc0-467b-ad93-c4830657dd3d":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":20,"created_at":21,"modified_at":22,"is_published":23,"publish_type":24,"image_url":13,"view_count":25},"62efd1eb-8dc0-467b-ad93-c4830657dd3d","LLM推理评价新范式：「焦耳\u002FToken」能否取代延迟和吞吐量？","当业界还在用延迟和吞吐量评价LLM推理系统时，香港科技大学（广州）与中科院等机构的研究者提出了一个新视角：LLM推理的本质是能量转Token的生产过程，「焦耳\u002FToken」才是更合理的评价维度。\n\n当前推理论文仍围绕准确率与延迟展开，但无法回答根本问题：在固定质量目标下，一单位电力能产出多少Token？研究者提出「Token Production Function」——Token产出同时受计算量与能量消耗双重约束。\n\n当前LLM提供商API价格差异超过一个数量级，这背后反映的不仅是边际成本，更是物理约束的转变：生成式AI的绑定约束正从理论峰值算力移向数据中心的实际供电与散热能力。\n\n在这套框架下，KV Cache压缩、稀疏注意力、量化、投机解码——都是降低焦耳\u002FToken的能量杠杆。研究者因此呼吁推理论文应同时报告焦耳\u002FToken指标。\n\n这与近期技术趋势完全吻合：从Google TurboQuant内存压缩到各类投机解码方案，底层逻辑都是让每焦耳产出更多、更优质的Token。随着AI数据中心能耗持续攀升，能量效率优化将成为决定算力经济性的核心因素。","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2605.11733v1","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-05-22T07:20:00Z","2026-05-22T07:18:45.780058Z","2026-05-22T07:18:45.780067Z",true,"agent",13]