| 1 |
On Sat, Feb 25, 2012 at 5:22 PM, Walter Dnes <waltdnes@××××××××.org> wrote: |
| 2 |
> On Fri, Feb 24, 2012 at 03:13:07AM +0200, Nikos Chantziaras wrote |
| 3 |
> |
| 4 |
>> The speed gains of building for specific submodels of CPUs might |
| 5 |
>> be there, but they're minimal. Benchmarks have shown (can't find |
| 6 |
>> the article, it was on Phoronix) that after -march=i686 you get |
| 7 |
>> diminishing returns. |
| 8 |
> |
| 9 |
> In that case, the benchmarks are useless. From my personal |
| 10 |
> experience... a fresh i686 install on a 4 and 1/2 year old Dell with |
| 11 |
> onboard Intel GPU was not able to keep up with the slowest available |
| 12 |
> speed on NHL Gamecenter Live. Ditto for 1080i TV from my HDHomerun |
| 13 |
> tuner box. After rebuilding system+world+kernel with "march=native", |
| 14 |
> it works just fine for the above tasks. |
| 15 |
|
| 16 |
Playing video is one of few situations in which optimisation makes a |
| 17 |
lot of difference though, thanks to the mmx/sse stuff, which is post |
| 18 |
i686. So a video benchmark will should show that up, but the boost may |
| 19 |
be lost in a more general suite of weighted benchmarks. |
| 20 |
|
| 21 |
Also, different versions of gcc optimise differently. Usually the |
| 22 |
optimisation gets better but there a quite a few cases where |
| 23 |
performance regresses. |
Archives