EduardoS (no@spam.com) on 4/22/12 wrote:
---------------------------
>anon (anon@anon.com) on 4/22/12 wrote:
>---------------------------
>>I did not understand any of what you wrote.
>
>I shouldn't have expected any better from you...
>
>>My reply you quote here was your rhetorical question about Sandy Bridge, asking
>>why it is "not the same" as Bulldozer, or something.
>
>The previous sentence was about Power7, so the "not the same" question was comparing
>Power7 to SB, since it is too much to expect you to understand the question I'm not expecting an answear...

I still don't know what you're asking. Yes, POWER7 has faster L1 cache than SB at same size and larger process. I did not claim otherwise. Your infantile rhetorical questions seem to be attempting to "catch me out" of a position I am not taking.

Just state up-front what you mean, rather than this stupidity.

>>You claimed BD's "memory and cache" subsystems are very good, and I pointed to
>>one glaring example of where it is not very good.
>
>You pointed to two specific metrics wich doesn't make it good or bad.

The cache subsystem on BD is bad.

They may have done it for size or power constraints, but that does not make it good after-the-fact. It will be a significant limiter to BD performance is being limited, in my opinion.

>
>>It shows that L1 misses are frequent and costly.
>
>It also depends on the behaviour of the core, if the core stall on every L1 misses
>frequent misses are very costly,

Gee, thanks, Einstein. Good contribution.

> even if you L2 latency is just 10 cycles, if the
>core continue working for many cycles after many L1 misses than the cost of each
>miss may not be so high even with a 20 cycle L2, also, cycles have different costs for different processors.
>
>>I really don't know what you're talking about.
>
>Expected.
>
>>In my opinion, the problem is that you are incoherent, combative, and refuse to acknowledge my argument.
>>
>>BD has a small L1 and a slow L2.
>
>And you refuse to acknowledge that this is a weak argument.
>
>You could try to argue better, but with so "I don't understand/know" in your last post I doubt you will be able to.

I do know that cache latency is among the most important problems in any modern CPU in most workloads, and that BD's caches have significant weakness there, compared with SB.