re: <<"All I assumed was a 20% increase in IPC">>
<John, you are focusing on a narrow range of processor's applications. SPECrates are not visible or sensible to general public, where the bulk of PC sales have been and always will be. 4- or 8-16-way servers bring relatively small volume of CPU sales,>
2 way servers are the lifeblood of the server market and always will be. The numbers on spec.org prove that in CFP2000_rate, Intel will be WAY behind the Barcelona. THEY ONLY HAVE A FREAKIN 19% LEAD WITH FOUR TIMES AS MANY FPU'S -- 114 vs. 96. That's an 8-way score vs. a 4-way score. Xeon doesn't have a prayer of a chance in CFP2000_rate.
SPECrates are not visible or sensible to general public,
You are correct, hardly any "general public applications" can utilize 4 cores. So whay's your point? CFP2000 rate and CFP2006 rate are designed to utilize modern multiprocessors efficiently. The fact that Intel does so poorly at these benchmarks only points to the fact that when games and graphics software are written to take advantage of 4 processors, Intel CPU's will be choking on their FSB, and AMD processors will not.
it is inconsequential to focus on little possible advantages the current Barcelona could theoretically bring.
I would estimate that 1/3 of AMD's CPU revenues were from the server market for most of 2006.
The major issue is that current Opterons are 50% behind Core2 in major, single-thread performance, as per most recent SPEC2006 (66% in INT, and 32% in FP).
Single thread performance is totally irrelevant in the server market and the desktop market, what on earth are you talking about?
66% in INT
Not even correct. Highest CINT2006 is a dual core C2 Extreme 2.93 GHz that got 18.5. A 2.6 GHz DC Opteron got 12.9. Even with the clock speed advantage and AMD penalized for registered DDR2, it's 43%.
20% or even 40% increase in IPC (which IPC do you/they meant BTW?) will not bring AMD any performance crown, with all associated disadvantages in marketplace.
In dual socket CINT2006_rate, a 10% increase in IPC over K8 core will give AMD the performance crown with a 2.3 GHz Barcelona vs. a 2.66 GHz Xeon.
4-way (2x 2.8 GHz) CINT2006_rate is already at 51.7 spec.org
Derate by 2300/2800 for lower clock speed (though reduction is never that much) -- 42.5
Increase by factor of 1.8 for doubling number of cores -- 76.4this is more than justifiable -- going from single to dual core on a 4-socket system goes
from 46.9 to (93.1, 97.6) on (HP DL585, Sun Blade result, Tyan result, sorry I couldn't find
two results with identical hardware, but using a factor of 1.8 instead of 2.0 seems quite
pessimistic on my part seeing as AMD's CINT2006_rate numbers scale almost linearly in
number of CPU's OK, now add my 10% for L3 cache and the long laundry list of core improvements, and we are up to 84.1.
Guess what a pair of the fastest available Clovertown's got in CINT2006_rate? 82.2
Q.E.D.
Petz |
