Re: [f-cpu] elbrus

["Rob Finch" <robfinch@sympatico.ca>]

It's a good article. As the article itself states most of what it says has
been known for a few years already. I'm not sure I agree with it's
prediction of future hardware / software scalability. Certainly, scheduling
instructions in a superscaler architecture is a problem, but the
architecture is able to make use of most of the available ILP. It may be
true that a superscaler architecture does not scale well beyond a low ILP;
however is it true that the ILP of general purpose apps will suddenly
magically increase in the future? As I understand it: 1) Many general
purpose algorithms have limited ILP by thier nature. Until we learn how to
implement these algorithms in a parallel fashion, the ILP will not change.
(In fact many apps are mandated to a certain level of ILP). No amount of
software or hardware scheduling can fix this problem. The ILP will probably
generally remain within bounds for which a superscaler architecture can
provide good performance (my own prediction). 2) for general purpose
applications the instruction level parallelism averages between two and
three instructions( ie, its low). Yes, it is possible to implement a zillion
functional units nowadays, but it does no good to have 25 functional units
when only three of them are likely to be active at any one time. 3) As I see
it, a program cannot execute faster than it can update the state of the
machine and fetch operands. Are there processors out there with say 32 read
ports and 16 write ports, in order to support issuing instructions to more
functional units and completing several instructions at once ? (I really am
curious). The other alternative is to build something like a barrel
processor supporting SMT, but this isn't any better than a multiprocessor.
I could go on...

Rob

----- Original Message -----
From: "Graham Seaman" <graham@seul.org>
To: <f-cpu@seul.org>
Sent: Monday, March 04, 2002 5:06 PM
Subject: [f-cpu] elbrus


> Anything interesting/rlevant in this article on the elbrus?
>
>
> http://www.rons.net.cn/english/FSM/english/FSM/ISSUE02/e2k.pdf
>
> Graham
>
> *************************************************************
> To unsubscribe, send an e-mail to majordomo@seul.org with
> unsubscribe f-cpu       in the body. http://f-cpu.seul.org/
>

*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/