Main Page | Report Page

 

  Computers Forum Index » Computer Architecture » mips per watt - ARM vs. X86...

Author Message
Paul A. Clayton...
Posted: Wed Oct 06, 2010 12:45 am
 
On Oct 2, 12:18 pm, n... at (no spam) cam.ac.uk wrote:
[snip]
Quote:
The most permanent reason for favouring fixed-width instructions
is that it makes diagnostics, debuggers and first-level interrupt
handlers rather easier to write and more reliable in practice.
That will remain true as long as systems are programmed in unsafe
languages, like assembler or C/C++.

Would prohibiting instructions crossing some moderate-sized
alignment point be helpful (without having to sacrifice much
of the code density advantage of variable width)? (Instruction
translation also becomes significantly easier.)

Even just alignment of procedure entry points might have
some benefit. (Is it ironic that higher-performance x86
use aligned instruction fetch--with highly variable length
instructions--, but take significant effort to support
unaligned data access? The MIPS R10000 could fetch
up to four instructions starting at any valid address--
admittedly without the ability to cross cache lines.)


Paul A. Clayton
just a technophile
 
Paul A. Clayton...
Posted: Wed Oct 06, 2010 1:09 am
 
On Oct 2, 6:05 pm, Brett Davis <gg... at (no spam) yahoo.com> wrote:
[snip]
Quote:
Fixed width made sense for about one process node, and
even that is a stretch. In short it was STUPID.

If Cortex or MicroMIPS had shipped back in the 1980's
instead of ARM or MIPS, we might be complaining about
the ARM/MIPS monopolies instead of Intels monopoly.

Only if they had also provided a crippled 8-bit
memory interface version for the IBM PC. :-/

(I have no clue, but I would not be surprised base on
what little I have read on the matter that the MIPS
and ARM developers were much more academics/engineers
than business executives. It might have taken them
too long to accept that quality does not sell [by
itself].)


Paul A. Clayton
just a technophile
 
Andy Glew...
Posted: Wed Oct 06, 2010 4:19 am
 
On 10/4/2010 10:38 PM, Brett Davis wrote:
Quote:
In article<boidnf6VutXF3zTRnZ2dnUVZ5hOdnZ2d at (no spam) giganews.com>,
Andy Glew<"newsgroup at comp-arch.net"> wrote:

On 10/2/2010 2:00 PM, Brett Davis wrote:
In article<5NGdnZHsUZE77TrRnZ2dnUVZ5sWdnZ2d at (no spam) giganews.com>,
Andy Glew<"newsgroup at comp-arch.net"> wrote:
This of course may turn out to be wrong headed. Why not expand out that
third port and do double Adds, etc?

(2) By "double Adds", I assume you mean cascaded Adds:
R1 := R2+R2
R4 := R1+R5

I'm all for that (and even deeper logic trees). Although I tend to
think of it as being done dynamically, e.g. separate instructions that
are scheduled together, rather than having a single instruction that
does two adds.

No, that also adds an extra output, which seems crazed to me,
as one can barely track all the inputs and outputs now.

Fine, don't use an extra output:

R1 := R2+R3+R4

but that's a new instruction.

Or just make it work with idioms of existing instructions.

R1 := R2+R3
R1 := R1+R4

Quote:
But you are crazy like a fox. ;)

This gives a more balanced in/out system.
Three 2->1 ALUs is six inputs, but only 3 outputs, so three values
have to come from somewhere else, and the register file only has 2 ports.
Two 3->2 ALUs is also six inputs, but four outputs/results.

An extra net instruction executed per cycle at a wiring cost of just
one more output port, and ALU.

Hard to schedule for though, and worst case is bad.

Sun did this in one of their SPARCs, although, done badly, it may
explain some of the poor results of their systems.

Myself, I would actually just do it on a regular two wide superscalar
pipeline:

R1 := R2+R3
R4 := R5+R6

and give that pipeline (2 ops, each with 2 inputs and 1 output)
the ability to handle dependent ops in the same cycle

R1 := R2+R3
R4 := R1+R5

This is pretty easy to schedule for in a 2-wide superscalar machine.

If you want to do it in 2 sets of 2 at a time

R1 := R2+R3
R4 := R1+R5
R1' := R2'+R3'
R4' := R1'+R5'

It is a bit more of a pain - although equivalent to having a bypass cluster.


The motivation: it's pretty easy to do, circuit wise. I.e. it is low
hanging fruit.



Quote:
You are fundamentally adding an ALU, so wires get longer, and clocks
can drop. (Same as my suggestion, but less so.)

Or, just adding a bypass to an existing set of superscalar ALUs. making
them work in a few more cases.




By the way, GPUs do this.
 
Brett Davis...
Posted: Wed Oct 06, 2010 10:15 am
 
In article <1jpw6qk.k5arhqqxalt4N%nospam at (no spam) ab-katrinedal.dk>,
nospam at (no spam) ab-katrinedal.dk (Niels Jørgen Kruse) wrote:

Quote:
Brett Davis <ggtgp at (no spam) yahoo.com> wrote:

I am just so used to tech firms booming and busting.
One surviving mainframe company.

You forgot Unisys.

88% services, most of the rest Xeon.
Two mainframe lines appear to still be available,
a 36 bit system, and a stack based processor.
http://en.wikipedia.org/wiki/Unisys
http://en.wikipedia.org/wiki/UNIVAC_1100/2200_series
http://en.wikipedia.org/wiki/Burroughs_large_systems

Quote:
No surviving Mini-frame companies.

HP. Mini costumers who followed the path of least resistance have been
herded to Itanium, eg. PDP-11 -> Vax -> Alpha -> Itanium.

No surviving mini-computer lines?

Except for the PDP-11, I think someone still makes a clone of it.
Wrong, Mentec Inc no longer exists.
http://fixunix.com/dec/98198-mentec-us-gone.html

Brett
 
Piotr Wyderski...
Posted: Wed Oct 06, 2010 12:49 pm
 
George Neuner wrote:

Quote:
Under strict interpretation, the DCMA prevents caching copies.

Fortunately, I am not covered by DCMA.
Anyway, does that act forbid sane implementations of Java and C#? :-)

Best regards
Piotr Wyderski
 
Paul Gotch...
Posted: Wed Oct 06, 2010 1:17 pm
 
Piotr Wyderski <peter.pan at (no spam) neverland.mil> wrote:
Quote:
Fortunately, I am not covered by DCMA.
Anyway, does that act forbid sane implementations of Java and C#? Smile

What the CLR does (C# is a language not a runtime or linkage standard)
is have things called AOT or Ahead of Time compiled blocks. The
original author ships their component with the CLR bytecode and the AOT
blocks compiled using expensive optimisation techniques. They own the
copyright on both so there is no problem.

I know that MS do ship large portions of the .NET Framework as
bytecode only and compile it into AOT blocks on the client computer
when it is installed. They can do this as they can grant permission as
they own the copyright...

So in general yes insane copyright law places sane implementations of
useful technology in ambiguous legal territory. This is almost
certainly going to get worse when the ACTA comes along.

This is not the first time this has happened and it won't be the last.
Unfortunately most ordinary citizens don't either understand or care
about most of the laws that passed in their name.

-p
--
Paul Gotch
--------------------------------------------------------------------
 
George Neuner...
Posted: Wed Oct 06, 2010 2:24 pm
 
On Wed, 06 Oct 2010 10:49:39 +0200, Piotr Wyderski
<peter.pan at (no spam) neverland.mil> wrote:

Quote:
George Neuner wrote:

Under strict interpretation, the DCMA prevents caching copies.

Fortunately, I am not covered by DCMA.

You are if any of your products end up in the US or its territories.
If you aren't, count your blessings!

Quote:
Anyway, does that act forbid sane implementations of Java and C#? Smile

I'm don't know what you mean by "sane" in this context. I'm not
certain that smiley is meant to be ironic.

Quote:
Best regards
Piotr Wyderski
George

 
Niels Jørgen Kruse...
Posted: Wed Oct 06, 2010 9:35 pm
 
Paul Gotch <paulg at (no spam) at-cantab-dot.net> wrote:

Quote:
So in general yes insane copyright law places sane implementations of
useful technology in ambiguous legal territory. This is almost
certainly going to get worse when the ACTA comes along.

ACTA seems to have been neutered.

<http://arstechnica.com/tech-policy/news/2010/10/near-final-acta-text-ar
rives-big-failure-for-us.ars>

--
Mvh./Regards, Niels Jørgen Kruse, Vanløse, Denmark
 
Bernd Felsche...
Posted: Thu Oct 07, 2010 5:15 am
 
Paul Gotch <paulg at (no spam) at-cantab-dot.net> wrote:
Quote:
Piotr Wyderski <peter.pan at (no spam) neverland.mil> wrote:
Fortunately, I am not covered by DCMA.
Anyway, does that act forbid sane implementations of Java and C#? :-)

What the CLR does (C# is a language not a runtime or linkage standard)
is have things called AOT or Ahead of Time compiled blocks. The
original author ships their component with the CLR bytecode and the AOT
blocks compiled using expensive optimisation techniques. They own the
copyright on both so there is no problem.

I know that MS do ship large portions of the .NET Framework as
bytecode only and compile it into AOT blocks on the client computer
when it is installed. They can do this as they can grant permission as
they own the copyright...

So in general yes insane copyright law places sane implementations of
useful technology in ambiguous legal territory. This is almost
certainly going to get worse when the ACTA comes along.

This is not the first time this has happened and it won't be the last.
Unfortunately most ordinary citizens don't either understand or care
about most of the laws that passed in their name.

It'd be good if the legislators voting on the bills before them,
individually understood the subject matter. One could require every
one to individually provide a short essay on the aims of the
legislation and how it fills or fails to meet those aims; or how it
exceeds the scope with unintended consequences. A short verbal quiz
before voting is needed before the vote to make sure that there's no
cheating!

The essay and the representative's actual vote would be recorded and
made publically available. The electorate then has an indication of
their representatives' competence and honesty.

The alternative for the representative is that they abstain from
voting on things that they don't understand.
--
/"\ Bernd Felsche - Innovative Reckoning, Perth, Western Australia
\ / ASCII ribbon campaign | For every complex problem there is an
X against HTML mail | answer that is clear, simple, and wrong.
/ \ and postings | --HL Mencken
 
Piotr Wyderski...
Posted: Thu Oct 07, 2010 10:59 am
 
Paul Gotch wrote:

Quote:
What the CLR does (C# is a language not a runtime or linkage standard)
is have things called AOT or Ahead of Time compiled blocks. The
original author ships their component with the CLR bytecode and the AOT
blocks compiled using expensive optimisation techniques. They own the
copyright on both so there is no problem.

I know that MS do ship large portions of the .NET Framework as
bytecode only and compile it into AOT blocks on the client computer
when it is installed. They can do this as they can grant permission as
they own the copyright...

Paul, thanks for your input -- I'm extremely surprised!
I didn't even dare to think that the worst enemy of the
compile-on-site technologies is the copyright law itself.

George: that's the implementation I consider to be sane
(optimization during installation would be even better).
Discarding the already optimized pieces of code and
repeating heroic optimizations of the same block over
and over again is technically just plain stupid vel insane.

Best regards
Piotr Wyderski
 
Ken Hagan...
Posted: Thu Oct 07, 2010 6:11 pm
 
On Thu, 07 Oct 2010 07:59:31 +0100, Piotr Wyderski <piotr.wyderski at (no spam) wp.pl>
wrote:

Quote:
Paul, thanks for your input -- I'm extremely surprised!
I didn't even dare to think that the worst enemy of the
compile-on-site technologies is the copyright law itself.

I think the cases Paul is thinking of are probably ones where the
copyright holder stood to lose by the form of copying being contemplated.
For example, if the vendor licences software for ARM and then you run it
on an x86-based ARM emulator, the vendor might take the view that they
could have sold you the x86 version. For JIT and compile-on-site, you are
running the product on the machine for which the vendor sold it to you, so
it would be hard for them to argue that they'd suffered any loss. This is
particularly true for Java or CLR applications where there must surely be
a presumption that you will run them on the standard "engines", which all
use this optimisation.

Since copyright is a civil matter, not a criminal one, if there is no loss
then the vendor would simply be wasting the court's time in bringing the
case. The court might take a dim view of that.

If Paul can point to cases where even JIT and compile-on-installation have
been found in breach of copyright, then I have to say that the court is
bonkers and we should not expect legislators to make copyright law even
more complex by adding text to permit the bleedin' obvious.
 
Rick Jones...
Posted: Thu Oct 07, 2010 9:29 pm
 
Bernd Felsche <berfel at (no spam) innovative.iinet.net.au> wrote:
Quote:
It'd be good if the legislators voting on the bills before them,
individually understood the subject matter.

Well, to a degree at least. Given the breadth of topics covered by
bills, I don't think that even a True Renaissance (Wo)Man could have
complete understanding of the subject matter of every bill that
crosses their path.

Quote:
One could require every one to individually provide a short essay on
the aims of the legislation and how it fills or fails to meet those
aims; or how it exceeds the scope with unintended consequences. A
short verbal quiz before voting is needed before the vote to make
sure that there's no cheating!

I thought I was quixotic... :)

rick jones
--
A: Because it fouls the order in which people normally read text.
Q: Why is top-posting such a bad thing?
A: Top-posting.
Q: What is the most annoying thing on usenet and in e-mail?
 
Philip Homburg...
Posted: Fri Oct 08, 2010 1:26 pm
 
In article <6njvn7xt61.ln2 at (no spam) innovative.iinet.net.au>,
Bernd Felsche <berfel at (no spam) innovative.iinet.net.au> wrote:
Quote:
It'd be good if the legislators voting on the bills before them,
individually understood the subject matter. One could require every
one to individually provide a short essay on the aims of the
legislation and how it fills or fails to meet those aims; or how it
exceeds the scope with unintended consequences. A short verbal quiz
before voting is needed before the vote to make sure that there's no
cheating!

I have a more cynical (realistic?) view on how the political process works.

The number one question when voting about any bill is whether the voters care
enough to effect re-election chances.

If, as usual, that is not the issue, then the MP will try to make the most of
his vote. Either by getting something in return from lobbyists or, far more
likely, by getting other MPs to vote for bills he cares about.

Only the MPs who actually care about a particular bill will be worried about
its contents.


--
That was it. Done. The faulty Monk was turned out into the desert where it
could believe what it liked, including the idea that it had been hard done
by. It was allowed to keep its horse, since horses were so cheap to make.
-- Douglas Adams in Dirk Gently's Holistic Detective Agency
 
Bernd Felsche...
Posted: Fri Oct 08, 2010 2:03 pm
 
nmm1 at (no spam) cam.ac.uk wrote:
Quote:
Rick Jones <rick.jones2 at (no spam) hp.com> wrote:
Bernd Felsche <berfel at (no spam) innovative.iinet.net.au> wrote:
It'd be good if the legislators voting on the bills before them,
individually understood the subject matter.

Well, to a degree at least. Given the breadth of topics covered by
bills, I don't think that even a True Renaissance (Wo)Man could have
complete understanding of the subject matter of every bill that
crosses their path.

Which has the impact of reducing the number of laws being passed. In
itself, a laudable goal.

Quote:
One could require every one to individually provide a short essay on
the aims of the legislation and how it fills or fails to meet those
aims; or how it exceeds the scope with unintended consequences. A
short verbal quiz before voting is needed before the vote to make
sure that there's no cheating!

I thought I was quixotic... :)

The ideas have merit, but I can see most bills being passed on the
casting vote of the chairman, with 0 votes for and 0 against.

Quorum should be set to require a minimum number of votes.

--
/"\ Bernd Felsche - Innovative Reckoning, Perth, Western Australia
\ / ASCII ribbon campaign | For every complex problem there is an
X against HTML mail | answer that is clear, simple, and wrong.
/ \ and postings | --HL Mencken
 
Tim McCaffrey...
Posted: Fri Oct 08, 2010 4:09 pm
 
In article <QLudnUYPULFrzjTRnZ2dnUVZ5gydnZ2d at (no spam) giganews.com>,
"newsgroupatcomp-arch.net" says...
Quote:


In the Andy Glew UIUC HaRRM (Hardware Register Renaming Mechanism) style
of OOO, you always allocate a new virtual register as the target for
every operation. (By the way, there may be earlier examples of this
technique, although I don't know of them; I am just saying "Andy Glew
UIUC HaRRRM" as a way of full identification, not as a claim to priority.)

Sounds like a hardware version of SSA (single static assignment).


- Tim
 
 
Page 5 of 9    Goto page Previous  1, 2, 3, 4, 5, 6, 7, 8, 9  Next
All times are GMT
The time now is Mon Sep 01, 2014 11:46 am