Speculative Threading - Intel shows new compiler tech at IDF.

[Guden Oden]

Anandtech has a page describing this, and what it essentially is (for those who don't immediately click the link to read the text there) is a technique to let the compiler mark sections of code that can be snipped out from a program that in of itself is difficult to parallelize, and then run these snippets on idle CPU cores/threads.

Unfortunately this technique requires hardware support also, so no existing CPU can use this technique, including our currently upcoming multicore games consoles...

It is interesting though, because it has the potential to not only make programmers' lives easier, it also allows for rather significant performance boosts, at least in the example given by intel. Maybe we'll see something like this exploited in the NEXT-next gen consoles.

 

[Inane_Dork]

Thanks for posting this.

I'm attributing all "automatic" threading to science fiction until further notice. Though, this one doesn't promise all that automatic a process. I would be interested in seeing what the compiler requirements are. If the compiler is finding snippets to thread on its own, eh... But if it's programmer guided, I could believe it.

 

[Carl B]

From reading this morning though, the impression I got was that it was really better suited to OOE architectures. It's still a very exciting development should it come to fruition, but this particular implementation and model doesn't seem as if it would be viable on the IOE cores of the console world.

 

[dskneo]

Thanks for posting this.

I'm attributing all "automatic" threading to science fiction until further notice. Though, this one doesn't promise all that automatic a process. I would be interested in seeing what the compiler requirements are. If the compiler is finding snippets to thread on its own, eh... But if it's programmer guided, I could believe it.

toshiba already demonstrated a working compiler in the decoding demo (the 48 SDTV steams).
It's in a article somewhere...

edit: i also remember kutaragi mention that the duck demo featured 16 real-time programs, but i think it wasn't a compiler....

either way, not science fiction...

 

[Carl B]

toshiba already demonstrated a working compiler in the decoding demo (the 48 SDTV steams).
It's in a article somewhere...

edit: i also remember kutaragi mention that the duck demo featured 16 real-time programs, but i think it wasn't a compiler....

either way, not science fiction...

Well from my understanding Toshiba's demo was more a demonstration of Cell's 'automatic qeueing,' not any sort of automatic threading. What this is about is taking normally single threaded code and automatically chopping it up (to a degree) in such a fashion that idle cores may be brought to bare on a situation. It's a sort of attempt at a coding/hardware 'holy grail,' as it were.

 

[dskneo]

Well from my understanding Toshiba's demo was more a demonstration of Cell's 'automatic qeueing,' not any sort of automatic threading. What this is about is taking normally single threaded code and automatically chopping it up (to a degree) in such a fashion that idle cores may be brought to bare on a situation. It's a sort of attempt at a coding/hardware 'holy grail,' as it were.

well my understanding is what toshiba said in the article... they used a compiler to multitask the job to multiple SPe's.
It happened in may or march....

 

[incurable]

We've heard about helper threads for quite some time now, but I guess it'll take a few more years to get it shipping.

On the other hand, it seems rather "netburstian" to me, as it essentially trades off energy consumption for processing speed. Personally, I see only limited applicability for this in the new age of Performance/Watt.

 

[Saem]

http://imdb.com/title/tt0434409/plotsummary

This has nothing to do with V the 80s series.