计算性能增速远超片外内存容量与带宽，后者约每3.3-3.5年才翻倍。自2010年以来，训练算力已增长约1万亿倍，预计到2028年有效算力将再提升1000倍。一旦算术能力超越内存与网络传输能力，峰值浮点运算不再是瓶颈，数据移动将成为制约智能计算的关键负担。2026年4月的一项集群研究显示，长期运行的GPU任务约有19%的时间和10.67%的能耗处于“执行空闲”状态，此时时钟保持高位但有效计算极少。
Compute performance rising much faster than off-chip memory size and bandwidth, which roughly double only every 3.3 to 3.5 years.
Training compute rose by about 1TX since 2010 and will rise another 1,000X in effective compute by 2028.
And that's a big problem, once arithmetic outruns memory and networking, peak FLOPs stop being the bottleneck and data movement becomes the tax on intelligence.
An April 2026 cluster study estimates long-running GPU jobs spend about 19% of time and 10.67% of energy in execution-idle, where clocks stay high while little useful work is happening.
Since I began work on AI in 2010, training compute for frontier models has grown by one trillion times. Now we're looking at something like another thousand-fold growth in effective compute by the end of 2028. 1000x the existing 1,000,000,000,000x. Extraordinary stuff.在 X 查看被引用的帖子
来源：Rohan Paul · x.com