Quote Originally Posted by terrace215 View Post
No "infos say K10 has 20-50-70% IPC boost".

Now all AMD will say is that FOUR K10 cores @ 2GHz will perform 40-70% faster than TWO top-bin K8 cores, in certain cherry-picked benchmarks.

That's not a statement about K10 core IPC vs K8 core IPC.

That figure is likely 10-15% in most cases, putting Barcelona IPC roughly on par with Core 2 Duo IPC.

In applications that max out main memory bandwidth, the K10 will do relatively better. In apps with a large cache footprint, C2D will do relatively better.

But enough is known about the K10 design, plus AMD's simulated performance on a few benchmarks, to pin down its IPC.

I can understand any confusion if you only looked at AMD press releases, since, when they issue new claims, they fail to mention that the old claims no longer apply. And they've changed their story at least 3 times at this point:

First: Barcelona will outperform Intel's fastest Clovertown by X & Y%.

(AMD begins to have clockspeed concerns, so in April: )

Second: Barcelona will outperform Intel's Clovertown clock-for-clock by X and Y% on two 2P spec_rate benchmarks.

(Intel improves its compiler and tuning on spec_rate, and Barcelona clock is still dreadfully low, so change to: )

Third: Barcelona will outperform AMD's own DC K8, per socket (not core) by X and Y% on two types of applications.

and also:

Third(b): Barcelona will have the best performance PER WATT.

(But now Intel will launch 50W 2GHz Clovertown w/1333FSB in August, so AMD will be abandoning 3b shortly. )
You are speculating and presenting it as Gospel. None of us knows the IPC of K10. It is impossible to get a clear picture from AMD's statements. You want to speculate thats fine, just dont present your speculations as fact. As for compiler tuning,why would you run a K10 benchmark on code produced by the Intel compiler? If you meant that you would use Intels compiler for Clovertown and something else for K10 thats a really bad benchmarking practice as the scores are not directly comparable.