Unverified50% confidenceFactExact time
对于大规模推理集群，能耗往往占TCO的30%-50%
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
AMD MI355X运行GLM5.2：推理成本仅为英伟达Blackwell的一半
hackernewshackernews7/3/2026
Related Claims
Unverified建议上下文占用接近30%时开始关注，超过40%时执行Compact，超过60%时考虑Clear后重新开始66% similarUnverified推理时的KV缓存、激活值缓冲区和中间计算状态会额外占用30%-50%显存66% similarUnverifiedIn large-scale clusters, All-to-All communication can account for 30%–50% of the total execution time of a MoE layer.61% similarUnverified评判CPU与GPU仿真的TCO需针对具体工作负载建模，若任务在GPU上并行效率仅30%而CPU上达90%，最终有效成本可能颠倒61% similarUnverifiedIn High mode, the token consumption per chain-of-thought reasoning call may be 10 to 50 times that of Instant mode.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/186798API
curl https://kongchang.com/api/v1/knowledge/claims/186798MCP
get_claim(id=186798)