[Arm-netbook] EOMA server standard

[luke.leighton]

On Fri, Oct 26, 2012 at 9:03 AM, Roman Mamedov <rm at romanrm.ru> wrote:

> So I believe that with EOMA, "1 Gbit ought to be enough for anyone." :)

 i'm slowly coming round to this as well, for exactly the reasons you
say - which is that low-cost, low-power parallelism should eat
high-cost, high-power serialism hands down.

 one major, major disadvantage that x86 esp. intel designs have is
that the ALUs tend to have *multiple* units doing the *exact* same
calculation.  the reason for this is to get that last ounce of speed,
speed, speed.  i heard for example that intel had *five* separate
teams independently design a "divide" unit.  rather than pick the best
one and put that in, they put *all* of them in because they had
different characteristics.  the first one that answered was the one
that was picked - all other 4 computations were cancelled.

 the problem with this approach is that the ALU takes *far* more power
than should actually really be needed, just to get that last little
bit of extra serial performance.  you cannot use those 5 independent
units to perform 5 divide operations in parallel, so of course the
performance-per-watt when compared to parallel division of labour
across single-unit ALUs is much greater.

l.