By: Linus Torvalds (torvalds.delete@this.linux-foundation.org), November 13, 2014 5:57 pm
Room: Moderated Discussions
Doug S (foo.delete@this.bar.bar) on November 12, 2014 8:41 pm wrote:
>
> Intel was forced to abandon GHz based marketing when they went back to the P6 core which had lower
> clock rates but higher performance when the P4 hit the power wall
Really, people complained about P4 on other reasons than power.
The thing is, P4 just didn't perform that well. It had some really bad fragile failure points, and while Intel ended up fixing the really obvious ones (early problems with 16-bit segments, nasty ALU-shifter imbalance, some others) it still had tons of nasty glass jaws. Exception handling was horrid, and any branchy code with bad prediction would simply not perform all that well.
People always talk about the "power wall", but P4 failures were not just about that. P4 made it possible for AMD to kick Intel ass, largely because P4 just wasn't very good. Not even at high frequencies. I don't think it is at all a coincidence that AMD did so well with Opteron back when Intel was pushing P4.
At Transmeta, we'd look at x86 performance (for some obvious reasons), and the difference between P4 and Pentium-M (Banias, Dothan) was night-and-day. Sure, for some tuned loads (particularly well-behaved kernels of code like media etc that had been scheduled for P4), the P4 could be pretty stunning. But generally nowhere else.
The whole "power wall" argument is simplistic. The P4 was bad for other reasons too. And yes, it would have looked better at twice the frequency, but dammit, so would anything else. That's not at all P4-specific.
Linus
>
> Intel was forced to abandon GHz based marketing when they went back to the P6 core which had lower
> clock rates but higher performance when the P4 hit the power wall
Really, people complained about P4 on other reasons than power.
The thing is, P4 just didn't perform that well. It had some really bad fragile failure points, and while Intel ended up fixing the really obvious ones (early problems with 16-bit segments, nasty ALU-shifter imbalance, some others) it still had tons of nasty glass jaws. Exception handling was horrid, and any branchy code with bad prediction would simply not perform all that well.
People always talk about the "power wall", but P4 failures were not just about that. P4 made it possible for AMD to kick Intel ass, largely because P4 just wasn't very good. Not even at high frequencies. I don't think it is at all a coincidence that AMD did so well with Opteron back when Intel was pushing P4.
At Transmeta, we'd look at x86 performance (for some obvious reasons), and the difference between P4 and Pentium-M (Banias, Dothan) was night-and-day. Sure, for some tuned loads (particularly well-behaved kernels of code like media etc that had been scheduled for P4), the P4 could be pretty stunning. But generally nowhere else.
The whole "power wall" argument is simplistic. The P4 was bad for other reasons too. And yes, it would have looked better at twice the frequency, but dammit, so would anything else. That's not at all P4-specific.
Linus
