A long read, but an interesting one...

[adz]

Haven't quite finished it yet, but it is quite fascinating :-)

See here.

(Added by Admin)

In the third part of our NetBurst Architecture investigation trilogy we are going to reveal the details of the Replay mechanism Implemented in Intel Pentium 4 processors, which Intel keeps quiet about. This particular mechanism and its working principles explain why Pentium 4 processors perform pretty slowly, despite their high working frequencies.

[Cymric]

Clever blokes, but they really need to learn that you do not need 17 pages to explain the end result. In addition, it would have been nice if they had not only demonstrated the problem, but showed the software engineers how to write proper code so Replay wouldn't kick in all the time. You can't have everything :-).

It just goes to show that you should not add speculative kludges to systems where everything depends on timely scheduling and execution. Unfortunately, I've got a nasty feeling that predicting execution time exactly is an NP-hard problem, so you have to make do with kludges to get anything going. However, they could have been made a bit more intelligent than Replay.

[bloodline]

Cymric wrote:
Clever blokes, but they really need to learn that you do not need 17 pages to explain the end result. In addition, it would have been nice if they had not only demonstrated the problem, but showed the software engineers how to write proper code so Replay wouldn't kick in all the time. You can't have everything :-).

No point, netbust is dead.

It just goes to show that you should not add speculative kludges to systems where everything depends on timely scheduling and execution. Unfortunately, I've got a nasty feeling that predicting execution time exactly is an NP-hard problem, so you have to make do with kludges to get anything going. However, they could have been made a bit more intelligent than Replay.

They could have... but then it would have been called the Pentium III-Extreme (Or Pentium-M as I like to call it) :-)

[Cymric]

bloodline wrote:
No point, netbust is dead.

Well then, why bother flogging a dead horse in such great detail? I think there are more Pentium 4's used around the globe than you care to admit, and these things don't run on their own.

[bloodline]

Cymric wrote:

bloodline wrote:
No point, netbust is dead.

Well then, why bother flogging a dead horse in such great detail? I think there are more Pentium 4's used around the globe than you care to admit, and these things don't run on their own.

Yeah, I mean it's a dead end... software that is running on the Netburst now is already achiving acceptable performance... there's no point worrying about it. Intel have stated quite clearly that they want to move away from it, with much noise in the direction of the Pentium-M (as coverted by apple), and IMHO a rather nice chip (good performance/power usage trade off).

[cecilia]

retard HW question:

does this mean my P4 can be faster?

[bloodline]

cecilia wrote:
retard HW question:

does this mean my P4 can be faster?

It means that programmers could optimise their code better for the Pentium 4, to avoid Replay loops, which delay execution.

um... in a word... yes.