By: NobodyatAll (None.delete@this.none.com), June 4, 2013 3:23 pm
Room: Moderated Discussions
Intel used the geometric mean of the benchmark scores, which is not only the correct mean in the case if rates comparisons, it is also considerably more conservative...Silvermont may actually still have an upside surprise in store on some benchmarks.
ARM presented to my knowledge very little information about how they were measuring performance in that chart. That community often resorts to DMIPS as a measure to make their chips look as good as possible against x86, so that wouldn't surprise here. Dhrystone, though, is not particularly meaningful in this use case, especially compared to the battery of tests Intel says they subjected Silvermont to.
Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on June 4, 2013 4:09 pm wrote:
> Exophase (exophase.delete@this.gmail.com) on June 4, 2013 10:36 am wrote:
> > anon (anon.delete@this.anon.com) on June 4, 2013 4:03 am wrote:
> > > Why would you think that? Intel has not proven to be particularly honest when they have not had
> > > the advantage -- which up to now has been "most of the time", when trying to compete with ARMs.
> > >
> >
> > Intel's tests comparing Silvermont and Saltwell are more likely to be accurate than tests between
> > an x86 and ARM processor. They don't have a strong incentive to make their current product look
> > bad when it'll be selling for another six months. So I doubt they deliberately sabotaged it by
> > using inferior binaries than they could have, using poorly representative test setups, etc.
>
> I wouldn't expect them to sabotage Saltwell (that would be crazy), but instead show Silvermont
> in the best possible light. For example show the best-case improvements rather than average
> improvements over a large range of benchmarks, and for iso-power/performance choose the point
> in the graph where the difference is maximal rather than the average over the full operating
> range. Simply showing the power/performance graphs for Saltwell and Silvermont would be far
> more informative and interesting but I'm sure there is a reason they weren't shown.
>
> > Most of Intel's test numbers were comparing with Saltwell, and those numbers alone
> > paint a strong picture. As for comparisons between Saltwell and various ARM cores,
> > I'd feel better using third party tests - fortunately those exist.
>
> I suppose from those ARM calculated the projected Silvermont results.
>
> Wilco
ARM presented to my knowledge very little information about how they were measuring performance in that chart. That community often resorts to DMIPS as a measure to make their chips look as good as possible against x86, so that wouldn't surprise here. Dhrystone, though, is not particularly meaningful in this use case, especially compared to the battery of tests Intel says they subjected Silvermont to.
Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on June 4, 2013 4:09 pm wrote:
> Exophase (exophase.delete@this.gmail.com) on June 4, 2013 10:36 am wrote:
> > anon (anon.delete@this.anon.com) on June 4, 2013 4:03 am wrote:
> > > Why would you think that? Intel has not proven to be particularly honest when they have not had
> > > the advantage -- which up to now has been "most of the time", when trying to compete with ARMs.
> > >
> >
> > Intel's tests comparing Silvermont and Saltwell are more likely to be accurate than tests between
> > an x86 and ARM processor. They don't have a strong incentive to make their current product look
> > bad when it'll be selling for another six months. So I doubt they deliberately sabotaged it by
> > using inferior binaries than they could have, using poorly representative test setups, etc.
>
> I wouldn't expect them to sabotage Saltwell (that would be crazy), but instead show Silvermont
> in the best possible light. For example show the best-case improvements rather than average
> improvements over a large range of benchmarks, and for iso-power/performance choose the point
> in the graph where the difference is maximal rather than the average over the full operating
> range. Simply showing the power/performance graphs for Saltwell and Silvermont would be far
> more informative and interesting but I'm sure there is a reason they weren't shown.
>
> > Most of Intel's test numbers were comparing with Saltwell, and those numbers alone
> > paint a strong picture. As for comparisons between Saltwell and various ARM cores,
> > I'd feel better using third party tests - fortunately those exist.
>
> I suppose from those ARM calculated the projected Silvermont results.
>
> Wilco