Xbit Labs found a major fault in the Netburst architecture?

[Jeff7181]

Link

Not sure I fully understand it, but it's interesting to say the least.

Summing up I can say that replay is not an independent trait of the NetBurst architecture, intended to increase the processors performance. Replay is more likely to be regarded as ?the other side to the picture? of the longer pipeline, as an auxiliary mechanism intended to help resolve speculation issues. The performance drop caused by replay is the price we have to pay for high working frequency. That is probably why we can rarely come across some scarce mention of the replay system in Intel?s official manuals and docs. Replay very often causes unjustified waste of resources and significant performance drops.

Since there is no description of replay causes and consequences in the official documentation, many software developers simply have no idea that it exists and thus cannot optimize their programs for it. It is definitely good news that Prescott processors acquired replay queues reducing the negative influence of the replay on two parallel working threads when HT technology is enabled, however, even in this case we see a 20% performance drop in the second thread cause by the replay of the first one.

 

[CheesePoofs]

Could someone summarize it for those of us who are way to lazy to read that much?

 

[Jeff7181]

Originally posted by: CheesePoofs
Could someone summarize it for those of us who are way to lazy to read that much?

Sometimes the last page of an article is the conclusion/summary... but if you're too lazy to click on the link, and then on the link to the last page I'll edit my OP with the summary.

 

[Kell]

"Major fault?" Hardly. It doesn't cause any errors or failures, unless the performance hit causes a real-time task to miss a deadline. And if you've got a real-time deadline to worry about, you probably shouldn't be using x86 anyways.

I'd consider this a performance quirk. Intel introduced some new tech that leaves an opening for the pipeline to get stuck. Intel apparently noticed this in testing and added extra logic to quietly get it unstuck when necessary. Sure, it's kind of a hack, and it causes a performance hit, but that performance hit is just one of many that have been showing up in benchmarks for ages.

 

[CheesePoofs]

Originally posted by: Jeff7181

Originally posted by: CheesePoofs
Could someone summarize it for those of us who are way to lazy to read that much?

Sometimes the last page of an article is the conclusion/summary... but if you're too lazy to click on the link, and then on the link to the last page I'll edit my OP with the summary.

I read that, but it still didn't make much sense. The conclusion was written for someone who had read the whole article, so I was hoping for someone who acutally had read the article to enlighten me a bit. The whole thing seems rather confusing though, so I probably wouldn't get it anyways.