AMD strictly moderated (AMD)

[wbmw]

Re: It is as high as 30% in some apps which I can dig out those AMD links

Your problem is that either do not have the brain power or the desire to think about it logically. The IMC reduces memory latency from the range of ~70ns (for Intel dual core parts with desktop memory controller) to the range of ~55ns (for AMD dual core parts), a savings in first-to-load usage of ~20%.

http://techreport.com/reviews/2006q2/core-duo/index.x?pg=4

Now that's assuming that all memory requests miss the cache, that they all have to go out to memory, and that all the requests are spaced out such that none of them are pipelined.

But the cache hit rate on a Pentium D is going to be in the 90s percentile, even without all of today's prefetching and locality tricks. And the reality is that memory transactions are pipelined, so you aren't waiting the ~70ns for a first-to-load transaction, but rather 2-4ns for a bus turn-around for the next transfer to occur.

All this substantially reduces the best case win of 20% to something more like 2-4%. So the earlier comment that the IMC is aiding K8 performance, rather than the micro-architecture, is plainly wrong.

Now with respect to the previous generation, even the Athlon XP had a first-to-load latency as low as 92ns, and compared to the best single core A64 latency of 43nm, it gives you ~50% latency reduction.

http://techreport.com/reviews/2003q3/athlon64/index.x?pg=7

But again, the Athlon XP had a 512KB cache with hit rate near enough to 90%, as well as a bus that supported pipelined transactions. So that 50% shrinks 5-10x in real applications that don't have sparse transfers that always miss cache. So while 20% may be supported on some apps with strange traffic patterns that even prefetching can't solve, the more likely benefit is in the range of 5-10% for the K8 vs. K7.

These are the cold facts, but unfortunately, you are without the class, honesty or integrity to admit it. As you put it, words cannot fully describe the extent of your deception and mendacity.