1. Intel聲稱Netburst嘅branch prediction係業界最勁原來唔係吹水
詳情請睇:
http://www.hkepc.com/bbs/viewthr ... &extra=page%3D5
2. Conroe 2MB & 4MB L2行SuperPi 嘅分別冇以前想像嘅大...用PrefMonitor搵到答案~
下圖係Northwood行SuperPi 32MB嘅L2 hit rate...
行32M每過幾秒會hit rate會短暫下跌, 但512K/1M就冇. 原因係就算運算資料大過cache容量, CPU/compiler識得將data預先放入L2, 令L2 hit rate可以相對保持高位.
Conroe嘅道理都係一樣, 就算只有2MB都可以靠prefetch data黎保持L2 hit rate.
(估唔到data prefetch嘅準確度都相當高下~)
