Prev: Branchless RLE decoding (Was Re: Delphi needs opengl gui :))
Next: Running Windows or Ubuntu on a multi-core CPU is there a way to get the O/S to use only 2 of the CPU's so that the other 2/4/6 CPU's are "free" for programs; in the sense ...
From: Robert Myers on 8 Aug 2010 16:51
This chart

http://www.coherentlogix.com/index.php?option=com_content&view=article&id=48&Itemid=67

made me wonder, as usual, how one would feed such a beast, but it also
made me wonder if the power consumed by processing (performance/
processor-watt) matters all that much any more, and in what contexts.

In general, I would expect the power budget for high-throughput
processing to be dominated by the cost of moving data and not the cost
of crunching it.

I assume that the answer would be signal processing applications,
about which I know very little in practice. Can someone offer
perspective?

This message has been posted to comp.arch as well as to the google
group high-bandwidth computing.

Robert
From: MitchAlsup on 10 Aug 2010 13:20
On Aug 8, 3:51 pm, Robert Myers <rbmyers...(a)gmail.com> wrote:
> made me wonder, as usual, how one would feed such a beast, but it also
> made me wonder if the power consumed by processing (performance/
> processor-watt) matters all that much any more, and in what contexts.

As a point of reference: the CDC 6600, CDC 7600, and CRAY 1, spent 1/2
of the power budget (and cabnet volume) in memory.

It was often said that a vector processor is simply a (multi-channel)
DMA device that just happens to mangle the bits on the way through.

But I don't remember if the buffers driving and receiving from the
memory are counted on the memory side of the processor side.

Power mattles some n desktop and frame arranged computing.
Power mater big time in battery devices--the growth side of computing
as seen by FAB utilization.

Mitch
From: Robert Myers on 10 Aug 2010 13:48
On Aug 10, 1:20 pm, MitchAlsup <MitchAl...(a)aol.com> wrote:

> As a point of reference: the CDC 6600, CDC 7600, and CRAY 1, spent 1/2
> of the power budget (and cabnet volume) in memory.
>
> It was often said that a vector processor is simply a (multi-channel)
> DMA device that just happens to mangle the bits on the way through.
>
> But I don't remember if the buffers driving and receiving from the
> memory are counted on the memory side of the processor side.
>
> Power mattles some n desktop and frame arranged computing.
> Power mater big time in battery devices--the growth side of computing
> as seen by FAB utilization.
>
I think I understand the general pressure to reduce power consumption.

I grew up with and still think in terms of the kinds of computers you
named, and I'm still interested in high-throughput computing. The
device that got me thinking advertised very low processor power
consumption, but the data for all that number crunching has to come
from somewhere and go somewhere.

There must be something like an Amdahl's law for power consumption,
and it has to apply, at some point, even to mobile devices. That is
to say, processor redesign can reduce power no further than the power
consumed by all other devices. For some, the power consumed by all
other devices might be dominated by spinning disks or a display. For
my purposes, "all other devices" might as well be cables, memory
drivers, memory, switches, and routers.

Recent advances in power management have been impressive. How far can
this go before it doesn't matter any more?

There are applications that involve lots of crunching relative to the
amount of data movement. Apparently, graphics processing is one such
application. Maybe signal processing is another--I was hoping for
some perspective.

My repeated (and now tiresome) complaint is that all problems are
being reduced to problems with lots of crunching and very little data
movement not because that's where the problems are but because that's
what people who build computers know how to do. It isn't my intent to
keep repeating that complaint, but I'm still looking for insight as to
where the constraints really are.

Robert.

 | 
Pages: 1
Prev: Branchless RLE decoding (Was Re: Delphi needs opengl gui :))
Next: Running Windows or Ubuntu on a multi-core CPU is there a way to get the O/S to use only 2 of the CPU's so that the other 2/4/6 CPU's are "free" for programs; in the sense ...