Yesterday we showed that according to two SPEC benchmarks, Barcelona is only 5-10% better than current Opterons.
Today, we’re going to see how current Opterons fare in these two benchmarks against current Xeons.
Obviously a significant difference, and not a gap that would be closed if the Xeon were running at 3GHz.
Obviously a very big gap, definitely “wiping the floor with Intel” territory. No wonder Intel rushed SSE4 into play.
Most reading this won’t be running multisocket systems, at best, they’ll be doing quadcore. Are the results essentially any different when it’s four vs. four?
It’s basically the same tale.
Two Important Conclusions
First, when AMD says that Barcelona is going to be so much better than Intel’s stuff, they’re basing it on a benchmark that shows its current stuff to be so much better than Intel, too.
What is really happening is that Barcelona is just extending a current lead a little, not creating a big gap. It’s not going from even or behind to 40% better; it’s more like going from 35% better to 40% better.
What AMD is saying is technically accurate, but it’s misleading in that many might think Barcelona represents a big leap in performance when it’s really a small hop.
Some may say that the comparison ought to be against Penryns, but there’s a far more fundamental problem here. What these spec benchmarks are saying is that, even after adjusting for clock speed, the current AMD K8s are better than the current C2Ds, a few percentage points better in integer, a lot more in floating point. Barcelona is just the cherry on top of the cake.
In other words, to believe that the AMD claims about Barcelona based on SPEC have general applicability, you also have to believe that current K8s are better, clock-for-clock, than C2Ds.
To put it mildly, that flies in the face of a mountain of tested documentation to the contrary for desktop applications and games.
Perhaps SPEC has some, maybe a lot of validity for certain type of server applications, but it seems to be a poor predictor of overall desktop performance.