http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3326&p=1

AMD is toast

Especially the L3 is very good:

Quote:

The L3 cache is quite possibly the most impressive, requiring only 39 cycles to access at 2.66GHz. The L3 cache is a very large 8MB cache, 4x the size of Phenom's L3, yet it can be accessed much faster. In our testing we found that Phenom's L3 cache takes a similar 43 cycles to access but at much lower clock speeds (2.0GHz). If we put these numbers into relative terms it takes 21.5 ns to get a request back from Phenom's L3 vs. 14.6 ns with Nehalem's - that's nearly 50% longer in Phenom.

Wonder what AMD did with their L3 ... :(
I know, Intel is already using 45nm, but I doubt that Shanghai will do any miracles :(

cheers

Opteron

Amazing! Intel is really upping the ante on CPU performance!

I'm not one of those early adopter types though, so I'll wait till Nehalem makes it to 32nm before I buy it.

For now, my 3.6ghz Yorkie will suffice.

According to Charlie and others, AMD did not believe that Core 2 was as fast as it was appearing in early benchmark "leaks".

I hope it turned them into believers, and I hope that they have an answer.

So long Hector and co. :(

I don't see what the huge bugaboo about L3 performance is here. At comparable clock speeds the L3 latency will be the same - there is no difference.

For reference, L2 latency for Barcelona is 9 cycles for 512kb vice 11 cycles for Nehalem.

Nehalem is a multithread monster though, that's for sure.

Nehalem is about what the original K10 was to be, multi cored with smt and imc. OTH nehalem is wider like C2D is wider than K8. Might this be an important differentor between original K10 and nehalem?

Actually they are the same; I'm citing core clocks when I quote L3 latency.

What remains to be seen is what the best case latency for Nehalem's L2 and L3 are. The problem with a lot of cache tests are that they're sort of flawed when it comes to determining that. For instance, Penryn has a variable latency L2.

I don't really understand what do you mean by "comparable". If we look at comparable core clocks we will get:
2.5GHz Phenom (L3 running at 2GHz): 21.5ns
2.66GHz Nehalem: 14.6ns

=> Nehalem's latency is clearly better.

Anand's measurements are correct but reported in the wrong context.

If the core sees the L3 in 43 cycles, you have to measure the absolute latency in nanoseconds at core clock frequency, and not northbridge clock frequency. The Northbridge clock could be running at 1 mhz, but if the core clock only sees it after 43 core clocks, you don't do the latency measurement using 1 mhz. Is that more of a clear explanation?

I am making the bold assumption, of course, that Anand ran the 2.5 ghz frequency part with a 2.0 ghz northbridge.

The correct number would then be 17.2 ns. at 2.66 ghz it would be 16.1 ns. Clearly, 39 < 43 cycles, but again, Barcelona's L3 latency is variable depending on numerous factors including NB clock, core clock, frequency of access, ratio of NB/Core clocks, bypass etc. Best case is < 43 cycles.

So in the context of the thread, jumping on AMD for cache performance is not the correct argument to make for saying why Intel is ahead in performance with the Nehalem part - in fact, you should wonder why the cache performance isn't better given Intel's historical cache advantage.

Here's a hint: Having 2 threads access a cache make the latency go longer.