Expired60% confidenceFactExact time
In a complete AI application system, GPU inference typically accounts for only 30%–40% of total compute, with the remaining 60%–70% of the compute load borne by the CPU.
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 9/30/2026(expired)
Sources
Related Claims
Unverified相较于用GPU执行AI推理，NPU的功耗通常只有前者的1/5到1/1077% similarUnverifiedWhen AI clusters scale to tens of thousands of GPUs, overall efficiency can plummet to below 40%, with more than half of computing power consumed in data transmission and waiting.76% similarUnverifiedAI GPU profit margins are as high as 70%–80%, far exceeding the 50%–60% of consumer-grade GPUs.75% similarUnverified单条推理通常只能占用GPU SM利用率的5%～15%，而合理的批次大小可将利用率提升至60%～90%75% similarUnverified许多推理服务的GPU利用率不足30%，通过VGPU技术可将利用率提升至70%以上74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/48910API
curl https://kongchang.com/api/v1/knowledge/claims/48910MCP
get_claim(id=48910)