Unverified50% confidenceFactExact time
H100内存子系统的动态功耗在大模型推理中占比可达整卡功耗的20%～35%
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
Verified思维链效果在参数量超过100B的模型上尤为显著73% similarUnverified在35B参数模型中GEMM内核往往占据总计算时间的60-80%72% similarUnverifiedAI推理成本主要构成是A100/H100等高端GPU的折旧与电力，属于资本密集型刚性支出，难以随收入下滑而快速压缩68% similarUnverified上下文控制在100K token以内时模型表现最为稳定，超过100K后幻觉率明显上升68% similarUnverified对于80%的日常编码任务，轻量模型的表现与旗舰模型差距在5%以内，但成本可能相差20-50倍67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/575667API
curl https://kongchang.com/api/v1/knowledge/claims/575667MCP
get_claim(id=575667)