Does anybody know benchmarks or head to head comparisons between CPUS (RISC o CISC) with Large L2 Cache againts Large L3 ones... (from 4 to 8 MB of cache size)????
In general terms L2 is fastest than L3 so which numbers or loads justifies the L3 rather than L2???
>why 4 MB L3 with 256KB L2 is better than 8 MB L2???
Says who ? It depends, it depends how fast you can make such a big L2 (and/or the L3), cache latency, associativity, inclusive or exclusive, etc. There is no easy answer. One is not necessarely "better" than the other, it depends on the implementation and so many factors, its just a design consideration.
= The views stated herein are my personal views, and not necessarily the views of my wife. =
It depends mostly on the program being run. If the program has inner loops that can fit into 256KB of L2 cache, it is better to have a fast, 256KB of L2, rather than a slow, 2MB of L2. However, the main program may fit into the 4MB of L3 cache. It depends wholely on the characteristics of the program you're running as well as the architectural properties of your processor. If your processor is low-clock speed and memory subsystem is low-latency, then having a large L3 is useless as your DRAM speed may be just as fast as your L3 cache anyway (think about Apple's backside bus with its DRAM used as L3 cache). On the other hand, if you have a P4 or Xeon with dramatic clockspeed gains vs a slow memory subsystem (relatively), you'd want several levels of cache. An L3 would be an improvement over main memory, and should be big enough to hold the majority of the running program and its data set. A fast L2 cache would be preferable and should be able to hold the inner loops of said program along with its current working data set. A fast L1 data cache should be able to hold a small portion of that data that's being access repeatedly. A small and extremely fast L1 instruction cache should be used so that the loop's instructions can be iterated over and over again in a very fast way.
"We are Microsoft, resistance is futile." - Bill Gates, 2015.