Unverified60% confidenceBenchmarkExact time
若带宽达到2.4TB/s，100GB模型理论速度上限可达24 token/s，60GB模型40 token/s，35GB模型接近69 token/s
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
9/23/2026
First Seen
Valid until: 12/22/2026
Sources
Related Entities
Related Claims
Unverified以70B的BF16模型为例，产生一个token需读入约140GB权重，打满H100的3.35TB/s显存带宽时每token延迟至少42毫秒，理论TPS上限约24 token/秒78% similarUnverified对于70B模型（序列长度2048、隐藏层维度8192），每层激活值约32MB，在InfiniBand(400Gbps)中约需0.6毫秒，而在50Mbps家庭宽带下需超过5秒78% similarUnverified在配备96GB显存的M2 Max上以1-bit精度运行时，简单任务生成速度稳定在约28-30 tokens/秒75% similarUnverified千兆网口全速传输理论速率约125MB/s，2.5G约300MB/s，万兆约1250MB/s，机械硬盘NAS(约150-200MB/s)配2.5G即可跑满，SSD阵列NAS才需万兆75% similarUnverified看SD卡标称速度时，UHS-I卡理论上限约104MB/s，即使标注U3/V30实际持续写入也难超90MB/s；需要更高持续速度必须选UHS-II卡（背面双排金手指）74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/945144API
curl https://kongchang.com/api/v1/knowledge/claims/945144MCP
get_claim(id=945144)