There was some speculation about AMD working on a mechanism that would
allow a multiprocessor/multicore system to gang-up the cores and use it
to execute single-threads much faster, also nicknamed as inverse
Hyperthreading. I think you could even call it "micro multithreading",
since it's finding threads at the atomic instruction level.

Here's a patent granted to AMD back in 2003, that seems to describe
this method. It's a bit of a boring read. From what I could garner in
the plain-talk summaries is that they are envisioning using each core
to speculatively execute an instruction stream, and so they execute the
same instruction stream in two cores simultaneously. If one of the
branches turns out right, then the results from the wrong branch are
discarded, etc. I'd rather have more qualified people read through it
and see what they think is really going on in here.

United States Patent: 6574725
http://patft.uspto.gov/netacgi/nph-P...&RS=PN/6574725

Yousuf Khan

Could be useful, who knows. Yet, then you'll have people complainin
and saying that AMD is taking Intel's ideas and using it fo
themselves. Yet, don't they all do that to eachother

Mike L wrote:
> Could be useful, who knows. Yet, then you'll have people complaining
> and saying that AMD is taking Intel's ideas and using it for
> themselves. Yet, don't they all do that to eachother?

Well, it's not something that Intel is doing right now. But then again,
neither is AMD, it's just got a patent, but no product based on it yet.
My only concern is that how much time is spent on speculative execution
of any single thread? How much time can you save doing this?

Another concern is if you're using both cores on a single thread, then
what happens when you do have multiple threads awaiting execution, you'd
be taking some power away from the true multithreaded apps while you're
servicing a single-threaded app.

Yousuf Khan

Yousuf Khan wrote:
> Mike L wrote:
> > Could be useful, who knows. Yet, then you'll have people complaining
> > and saying that AMD is taking Intel's ideas and using it for
> > themselves. Yet, don't they all do that to eachother?

Yes, except Intel has more money and hence does more R&D.

> Well, it's not something that Intel is doing right now.

Speculative multithreading...google is your friend.

> But then again,
> neither is AMD, it's just got a patent, but no product based on it yet.
> My only concern is that how much time is spent on speculative execution
> of any single thread? How much time can you save doing this?

Google is your friend.

> Another concern is if you're using both cores on a single thread, then
> what happens when you do have multiple threads awaiting execution, you'd
> be taking some power away from the true multithreaded apps while you're
> servicing a single-threaded app.

That is an issue, this wouldn't ever really work for servers.

DK

In article <449937cc$(E-Mail Removed)>, Yousuf Khan <(E-Mail Removed)> wrote:
>Another concern is if you're using both cores on a single thread, then
>what happens when you do have multiple threads awaiting execution, you'd
>be taking some power away from the true multithreaded apps while you're
>servicing a single-threaded app.

If you have more threads than cores then you wouldn't be using this. But
if you have cores going idle this might turn out to be a useful way to
put them to work. Even if the gains are minor, they are still gains.

This sounds like it gets more useful when run of the mill desktops start
having 4 or 8 cores. Using one core to speed up another by a few percent
doesn't sounds all that interesting now but as cores get cheaper it starts
sounding better and better. You have to spend the transistors somewhere.
--
Jim Prescott - Computing and Networking Group (E-Mail Removed)
School of Engineering and Applied Sciences, University of Rochester, NY

David Kanter wrote:

>Yes, except Intel has more money and hence does more R&D.

Much of which they flush on things like Itanic.

Jim Prescott wrote:

>This sounds like it gets more useful when run of the mill desktops start
>having 4 or 8 cores.

Gee, thanks for making everyone has has a dual-core CPU feel like
they're already behind the curve. 8)

*Not to mention those of us with (horrors!) only a single CPU in our
machines!

On a sunny day (Thu, 22 Jun 2006 10:05:22 -0500) it happened chrisv
<(E-Mail Removed)> wrote in <(E-Mail Removed)>:

>Jim Prescott wrote:
>
>>This sounds like it gets more useful when run of the mill desktops start
>>having 4 or 8 cores.
>
>Gee, thanks for making everyone has has a dual-core CPU feel like
>they're already behind the curve. 8)

Just buy a Sony PS3, it has a cell processor, there is your 8 cores ;-)

chrisv wrote:
> David Kanter wrote:
>
> >Yes, except Intel has more money and hence does more R&D.
>
> Much of which they flush on things like Itanic.

Perhaps it would behoove you to compare all the caches for different
CPUs...look at the latency and the power dissipation. You'll find that
Montecito's L3 is absolutely amazing. 0.75W/MB, nobody can come close
to that for on-die SRAM, especially with a 7.5ns (or less) latency.

Hardly a waste of money.

DK

> If you have more threads than cores then you wouldn't be using this. But
> if you have cores going idle this might turn out to be a useful way to
> put them to work. Even if the gains are minor, they are still gains.

There won't be gains and it won't be implemented. You need to have a
shared cache to get much benefits at all.

DK