AMD K10 to use Anti-Hyperthreading?

[pjbliverpool]

http://www.x86-secret.com/?option=newsd&nid=933

In English:

http://translate.google.com/translate?u=http%3A%2F%2Fwww.x86-secret.com%2F%3Foption%3Dnewsd%26nid%3D933&langpair=fr%7Cen&hl=en&ie=UTF-8&oe=UTF-8&prev=%2Flanguage_tools

If this is true then the potential performance increase could be huge, at least for single threaded apps. But perhaps even multithreaded apps will benefit if they are performance limited by individual threads.

I do wonder though how devs will aproach this as wouldn't it go in opposition to the current trend of writing apps with multithreading in mind?

Edit: Correct link added

 

[_xxx_]

Wrong link?

 

[pjbliverpool]

Wrong link?

Oops. sorry about that. Fixed now.

 

[Geo]

Well "SLI" and "CrossFire" are taken, so I wonder what they'll call it.

 

[fellix]

Anti-threading... sounds a bit odd to me. It's more like a "vector-ordered computing", regarding the thread managment in SMP/SMT enviornment (I think some super-computer mainframes had a similar capability).
Imagine a scalar/vector data-type processing in the modern CPUs (SSE & etc.), scaled up to a muti-threading level.

 

[Blazkowicz]

with a K8L rumored to have double the FP units (so, six FPU per core?) I thought AMD would go SMT.
I didn't know the K9 was cancelled (as said by first comment)

now I can't figure out if this "hardware threading" stuff makes sense.

 

[Blazkowicz]

wow, the translation is quite good
biggest error is, "More need then for the developers to spend the hours in the optimization of the software" instead of something like "no need anymore", or "no longer need" (is "no more need" okay?). in French we use the world "plus" for both meanings, auto-translation didn't get it.

 

[Bouncing Zabaglione Bros.]

Sun has something like this in it's Sunfire computers. You have whatever number of CPUs/memory in your box, and then you allocate them to virtual machines. You can even change the allocation of resources on the fly, or have all the virtual machines talking to each other like they are a networked cluster.

It makes a lot more sense to me that instead of having distinct threads for multiprocessing
on a number of cores/CPUs, you simply have a pool of resources to which you assign a program, and the multiple CPUs/cores seamlessly supply all the power they have to the program(s) being run on that pool. No messing with affinity, there's just a certain amount of processing power transparently supplied.

I wonder if this approach will work better for tasks that are not easy to multi-thread?

 

[3dilettante]

I'm trying to picture how this would work at a low level.
The translation doesn't make it sound like clustered multi-threading, which AMD has discussed before.

The scheduling hardware and register sets in each individual core currently cannot communicate to one another. Assigning a single thread to all cores implies that they have some means of knowing when there is a dependency or an exception occuring outside of their limited view. Without some kind of software or hardware mechanism, the cores are more likely to screw it up or wind up stalling every time they have to forward a result across a system request queue.

edit:

On second thought, this could be a from of clustered multithreading, though I wonder if such a chip would have multiple cores that are completely separate as they are now.

 

[NocturnDragon]

This sounds a lot like what intel was ready to show off with the canned Tejas processor: "Extended Enhanced HyperThreading".

It looks like Intel keept also going on in that way with the Mitosis Project so far getting a 2.2 performance increase on a single threaded Olden benchmark ran on a 4 processors system.

 

[Gubbi]

Sounds like bollocks to me.

At the very best it is something like helper threads, where one thread speculates wildly (producing incorrect results) to speed ahead and preload the caches with data the other thread then uses to produce correct results. Limited benefits and lots of caveats in that scheme.

Cheers

 

[NocturnDragon]

Btw that mitosis project, won't be ready before at least 5 years, and includes both hardware and software.

 

[ANova]

This sounds a lot like what intel was ready to show off with the canned Tejas processor: "Extended Enhanced HyperThreading".

It looks like Intel keept also going on in that way with the Mitosis Project so far getting a 2.2 performance increase on a single threaded Olden benchmark ran on a 4 processors system.

Yes, Intel has been working on this for awhile.