On Sep 30, 5:08=A0pm, chrisv <chr...@[EMAIL PROTECTED]
> wrote:
> aku ankka wrote:
> >A real genious can make even idiot understand what he means; he
> >masters the concepts he is conveying so well that he is able to
> >express them with least amount of effort. Since we all are human
> >beings with, let's assume, without any major defects in our brain we
> >should be on a fairly level footing here. I doubt that you have IQ
> >double of christv's, that would be highly improbable. So the chasm in
> >communication that exist is your own creation. Fix it
>
> You don't think that obfuscation is his goal? =A0I think it is. =A0Like
s=
o
> many with a losing position in the debate, he wants to twist the
> subject to an area that he thinks he can win. =A0Notice how he's managed
> to change the discussion from his illogic and his ****lling for the
> Intel Corp, to who has the bigger vocabulary? =A0
>
> Unfortunately for him, this group is about chips.
Intel Schintel AMD SCMD.. if we talk about chips, it's interesting to
note what (performance) problems those companies are trying to solve.
There are basically two kinds of bottlenecks for a single application
performance.
1. Volume of Data
2. Complexity
1. Is addressable with more functional units, ALUs, cores, ht's and
doesn't require high frequencies
2. Is more difficult problem to tackle, since the problems usually
faced in this domain are not parallelizable.. think of database query,
if the data is hierachical, the hierarchy must be traversed and this
is difficult to atomize into multiple processes .. problem isn't
solved by process technology, but reverting to case 1 (Volume of
Data) ; if possible, do multiple queries simultaneously, but this
isn't practical unless there are multiple clients doing queries
simultaneously, sometimes this is difficult to arrange.
Intel especially has lately been solving the (1) problem. The (2)
requires higher IPC and clock frequencies for a single state
machine .. this roadmap is more difficult than the other one. Only if
the software engineers would be so smart and See The Light and start
doing things in parallel intead of serially! But the ecosystem that is
x86 has too much inertia. It's good enough. It works well enough. It
makes enough revenue. It's proven; everyone "knows" how things should
be done. Programming languages like C/C++ are a plague, they are too
damn successful and that is a damn shame.
Low-frequency cores use less power. With same power profile, what
would you take: 2 x 3000 MHz parts, of 40 150 MHz parts? For the
current software ecosystem, you would take the 2 x 3000 MHz and so
would I. But in low-power profile environments, where the problems are
solved with custom software the 40 x 150 MHz is attractive. Requires
less power and probably much better performance overall, especially if
the ISA is crafted to a specific kind of tasks and isn't a do-it-all
like the x86. Graphics Processors are a good example, they are
basically a big array of multiply-ac***ulators and a scheduler
(command processor, what not..) and arbiters to keep the data on the
right path. Throw in some amount of fixed logic, some special purpose
functional units and you got the game covered.
Khronos Group is furiously working on a standard API for harnessing
these emerging capabilities. nVidia has some products out using their
proprietary API for the same task. AMD has their Fusion plans and
Intel recently introduced their new Architecture for similar purposes.
The sequential processing "model" of x86 is beaten to death and it
just keeps going.. it's success is the worst blocker for innovation
for everyone concerned.
|
