Re: From your chosen AMD platform, SPECint_rate = (41.4,56.0) for (2GHz,3GHz). That is 71% scaling, compared to 52% I showed for Clovertown, picking the very best scores available with an identical platform and testing date.
Which isn't surprising, actually, since the benchmark is now becoming memory bandwidth limited. In Clovertown's case, there are double the number of cores, but the same buses and bandwidths. To illustrate this point, I compared Woodcrest, which does have equal frequency scaling to Opteron in SPECint_rate. It just so happens that Barcelona *also* doubles the number of cores per socket without increasing memory bandwidth. So I would naturally expect it to have the same scaling issues. Of course, I'm not positive by any stretch, but it's better than you falsely proclaiming that Intel has frequency scaling issues without comparing apples to apples.
Re: You bring up the point that Intel dual-core C2D scales slightly better than quad-core [though still below AMD dual-core]. Therefore, you assume that Barcelona will scale more poorly than DC Opteron. However, there are no core changes between Woodcrest and Clovertown, so this is natural. In AMD's case, there ARE significant core improvements, so I would expect significant improvement in scalability.
Which features in particular are you expecting to affect frequency scaling? I would think that it's a memory bandwidth limited issue, and most of Barcelona's improvements are aimed at improving the base score (not SPEC_base, but the score at any given frequency). The improvements are not aimed as far as I can see at increasing, for example, the delta between 2.0GHz and 3.0GHz. |
