Re: Beyond multicore

Derek Simmons wrote:
You're probably right. Somewhere between 16 to 32 cores (or CPUs)
where going to hit a wall where the engineers are going to realize
that we need to change the way we design computers. The way computers
are designed date back to 1950 and 1960s because of cost factors.
Processor and memory subsystems were developed separately because they
were and sometimes still are expensive. Hence the model that we
adopted, the von Neumann Architecture. The problem we are running into
with the von Neumann Architecture is the amount of available bandwidth
between the memory and processing elements.

The last couple of computers I have upgraded/replace, when upgrading
the processor I was forced or to take full advantage of the new
processor I also needed to upgrade the RAM. So, if every time I
upgrade the processor I need to upgrade the RAM it got me to thinking.

What would be nice is if some large companies started competing to
design a new architecture that would combine the processor and memory
units like in the Connection Machine. It kind of makes sense
especially when you look picture of processor dies and about a third
to half the chip area is now cache RAM. The next step would be
designing a high speed interconnect that would allow routing and
communication between processors to be efficient. Proposals have been
made in the past but when MPP fell out of favor for cluster computing,
the funding for research for MPP architectures that could have benefit
multi-core/multiprocessor designs ended.

I imagine a machine that could be the cross breading of a Connection
Machine and SGI Origin2000 but smaller, something I could fit under my
desk. I could plug in modules that were a combination of memory and
processors to expand my computer as easily as I plug in memory DIMMs
now.

If such a machine were to be built then the next problem, software.
Right now there is a reluctance to switch to Vista because there are a
lot of known problems that existing software won't run under Vista. If
a new machine with a radical new architecture were to be designed then
new compilers to take advantage of it would need to be written (only
logical). A new OS kernel and virtual machines to support previous
operating systems.

When you think about it you begin to realize why nobody has undertaken
it - nobody wants to foot the bill. Maybe the government will step in,
like they did with HDTV and jump start it.


On Mar 7, 12:30 pm, Thomas Lindgren <***********@*****.***> wrote:

A bit of speculation here. I would expect that the performance gains
from multicore start tapering off at 8-32 cores per chip (or even
sooner). That would mean we will be running out of steam fairly
quickly (5 years or so?). So, my question is, what will be next?

My best guess at this point is going system-on-a-chip, exemplified by
Cell, Xbox360, or merging the GPU with a few CPUs. Revisiting on-chip
RAM might be promising too. But what more interesting options are
there?

(Can someone _please_ come up with an effective way to resume clock
scaling? :-)

Best,
Thomas
--
Thomas Lindgren

"Ever tried. Ever failed. No matter. Fail again. Fail better."




Yes, like the government is so adept at defining technology direction.

Actually you pretty much described Blue Gene, at least at a conceptual level.

The problem with this as a desktop/side model is there would have to be customers and software.

If you have a plan for those, the sandhill road boys would take care of you.

--
Del Cecchi
"This post is my own and doesn’t necessarily represent IBM’s positions, strategies or opinions.”
.

Relevant Pages

  • Re: Beyond multicore
    ... that we need to change the way we design computers. ... Processor and memory subsystems were developed separately because they ... adopted, the von Neumann Architecture. ... upgrade the processor I need to upgrade the RAM it got me to thinking. ...
    (comp.arch)
  • Re: Microsoft Zero Day security holes being exploited
    ... MS has flirted with *NIX in the past, from Zenix in the days of MS-DOS ... If anything, the upgrade cycle of NT is more forgiving than Win9x, ... in fact it's worse than that; they design purely for pro ...
    (microsoft.public.security)
  • Re: Antenna computor programs and pitfalls
    ... There's a lot of judgement and skill in antenna design, particularly when it comes to things like mechanical/electrical tradeoffs and manufacturability. ... There are tools designed to address one niche or another (e.g. there's programs that are designed to optimize electrical performance microstrip patch arrays, there's programs that are designed to optimize Yagi-Udas, etc.) ... No surprise that as computers came to be more common, such models would be first ones to be implemented. ... If for no other reason than you can compare the output of the modeling code with something you've actually built and see if it matches. ...
    (rec.radio.amateur.antenna)
  • Re: {OT:} You rightwingnuts need jobs and hobbies
    ... Wait, don't you play Bass? ... I design the cards we hand out with the demo package. ... That was when they were hungry for any computers they could get. ... One of these days, if I get a minute, I'll take the half a Corolla in the ...
    (alt.autos.toyota)
  • Re: Nokia set-top box: example of *dreadful* embedded design
    ... Reduced parts and bad design partitioning. ... >It's always been flaky, but since upgrading the firmware recently, it's ... saving config onto PC and restoring config after upgrade. ... A classic case of this was a router that on firmware upgrade added the ...
    (comp.arch.embedded)