AMD still in the game witih AM2

[stormy]

It's still whack, Vista or no Vista, 64 or no 64. Multiple CPUs running the same thread, given how most of today's processors work, is unfeasable.

This is new, old processors didn't do it because they couldn't. It's entirely possible that AMD has a new solution to the problem. A single core processor today is already superscalar, and assigns instructions to multiple data paths. If you have the assignment mechanism integrated across cores you could have one core stealing data paths from another and take over its resources. That is just me saying the first thing that came into my head. I'm really interested in what they did though. It is hard to extract more parallelism from one steam, this limit is what spurred the push to multithreading and dual cores in the first place.

[andyrose]

stormy, do you think dual processor platforms(with dual core processors) are going to catch on like SLI did?

[stormy]

stormy, do you think dual processor platforms(with dual core processors) are going to catch on like SLI did?

For normal users no, 99% of people just don't need it. By the time it could be usefull to the home market chip makers will have the additional cores offered on a single chip (4,8,+). I don't see home users needs outpacing that growth. But in industry multi-processor systems are already fairly standard (with many more than 2 chips) so it's still attainable for the people that like to blow wads on thier system.

[andyrose]

This would give you a huge boost though if you were using for, lets say, an audio studio with Sonar 4 or a video editing company, right? Is the processing boost going to be as big as the performance was on SLI videocards?

I'm asking you because you seem to know a lot about it, and I'm too lazy to look it up. Thanks.

[stormy]

The boost depends on how the programs are coded, which i don't know anything about specifically. If they are coded to run multi-threaded, and segment themselves well they will see a nice boost. I think that's easy for video processing, just encapsulate a bunch of frames and send those bundles off to different processors. Audio is probably similiar. Still, it comes down down to programming in regards to how well it works. Ideally you could approach a 1:1 relationship, double the cores halve the time, but this degrades as you add cores.

[gamevet]

The Saturn's dual processors were somewhat like a dual core, but the biggest hurdle is the bus.

[outRider]

This is new, old processors didn't do it because they couldn't. It's entirely possible that AMD has a new solution to the problem. A single core processor today is already superscalar, and assigns instructions to multiple data paths. If you have the assignment mechanism integrated across cores you could have one core stealing data paths from another and take over its resources. That is just me saying the first thing that came into my head. I'm really interested in what they did though. It is hard to extract more parallelism from one steam, this limit is what spurred the push to multithreading and dual cores in the first place.

Yes, not only are they superscalar, but also out-of-order, so already you've got one processor extracting the maximum possible parallelism out of a single thread, so what's left to do for other processors? The only time a single processor won't execute otherwise independent instructions in parallel is when they both depend on the same execution unit, and that doesn't happen too often, since optimizing compilers today try to interleave instructions as much as possible to avoid it.

But aside from that, it's still not feasable, because it would require the processors share a register file (at which point do you really have multiple processors?) or maintain seperate but synchronized register files, and a coherent cache hierarchy, which is just unrealistic, slow, and expensive.

[kedawa]

...people that like to blow wads on thier system.

LOL

[stormy]

Yes, not only are they superscalar, but also out-of-order, so already you've got one processor extracting the maximum possible parallelism out of a single thread, so what's left to do for other processors? The only time a single processor won't execute otherwise independent instructions in parallel is when they both depend on the same execution unit, and that doesn't happen too often, since optimizing compilers today try to interleave instructions as much as possible to avoid it.

But aside from that, it's still not feasable, because it would require the processors share a register file (at which point do you really have multiple processors?) or maintain seperate but synchronized register files, and a coherent cache hierarchy, which is just unrealistic, slow, and expensive.

A lot of things aren't feasable till somebody does it for the first time. The x86 ISA is over 25 years old, I wonder what things people thought were not feasable with it and the architecture as it evolved in that time. This news is only rumor, but there is still tons of research going on trying to increase parallelism so i wouldn't doubt that there will be more innovations someday.