By: Richard Cownie (tich.delete@this.pobox.com), November 17, 2010 10:55 am
Room: Moderated Discussions
someone (someone@somewhere.com) on 11/17/10 wrote:
---------------------------
>The McKinley core was a perfectly sane design for its
>era - large die 180 nm single core FSB based MPU.
Am I hallucinating, or wasn't the first IPF core called
Merced ? Has that one been airbrushed out of history ?
Anyway "perfectly sane for its era" depends on which era
you're talking about: the time they started designing it,
maybe around 93-94 ? Or the time it was first released,
in 2001 ? In 93 it was reasonable to think that
wide-issue in-order architectures might be competitive;
after PentiumPro (released late 95) and its derivatives
had achieved stunning success over the next couple of years
and kicked the ass of most of the RISCs, it was pretty
clear that the main technical assumption behind EPIC
(the idea that OoO was just too damn hard) was wrong.
From my perspective, the delays and the poor execution
look more like a symptom of having the wrong basic
architecture, rather than the cause of the failure.
They kept having to delay it because it was too damn
hard to make an EPIC core that could compete. And
every time they delayed it, the target moved as well.
But it doesn't really matter: either EPIC was a lousy
idea which couldn't be executed, or else it was at best
a mediocre idea which didn't offer sufficient advantage
to overcome a few missteps in management in execution.
You can have it either way. It certainly wasn't the silver
bullet they hoped for back in 1994 (and again in 2001,
and in 2003, etc ...)
---------------------------
>The McKinley core was a perfectly sane design for its
>era - large die 180 nm single core FSB based MPU.
Am I hallucinating, or wasn't the first IPF core called
Merced ? Has that one been airbrushed out of history ?
Anyway "perfectly sane for its era" depends on which era
you're talking about: the time they started designing it,
maybe around 93-94 ? Or the time it was first released,
in 2001 ? In 93 it was reasonable to think that
wide-issue in-order architectures might be competitive;
after PentiumPro (released late 95) and its derivatives
had achieved stunning success over the next couple of years
and kicked the ass of most of the RISCs, it was pretty
clear that the main technical assumption behind EPIC
(the idea that OoO was just too damn hard) was wrong.
From my perspective, the delays and the poor execution
look more like a symptom of having the wrong basic
architecture, rather than the cause of the failure.
They kept having to delay it because it was too damn
hard to make an EPIC core that could compete. And
every time they delayed it, the target moved as well.
But it doesn't really matter: either EPIC was a lousy
idea which couldn't be executed, or else it was at best
a mediocre idea which didn't offer sufficient advantage
to overcome a few missteps in management in execution.
You can have it either way. It certainly wasn't the silver
bullet they hoped for back in 1994 (and again in 2001,
and in 2003, etc ...)