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
VerifiedThe chain-of-thought effect is particularly pronounced in models with parameter counts exceeding 100B.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% similarUnverifiedThe main components of AI inference costs are depreciation and electricity for high-end GPUs such as the A100/H100, which are capital-intensive rigid expenditures that are difficult to compress quickly as revenue declines.68% similarUnverifiedModel performance is most stable when the context is kept within 100K tokens, and the hallucination rate rises noticeably beyond 100K.68% similarUnverifiedFor 80% of everyday coding tasks, the performance gap between lightweight models and flagship models is within 5%, but the cost difference can be 20–50 times.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)