By: Rumi Naik (easy.to.delete@this.guess.anyway), May 6, 2012 7:38 pm
Room: Moderated Discussions
Linus is completely right on this; does anyone remember the era when people wrote papers claiming IPC < 3 was the hard limit? The same phenomenon is at play: until use models challenge architectural needs, hard limits (while false) are essentially impossible to exceed. This isn't likely a matter of architectural scaling (see Amdahl's Law) but an issue of what the architecture is being asked to do in the first place.
I hope to see the day when Yale Patt's thesis on the systemic design of computing is taken more directly and at face value.
Linus Torvalds (torvalds@linux-foundation.org) on 5/5/12 wrote:
---------------------------
>>
>>According to the above results, the A9 has 31% lower IPC than Bobcat, ie. the improvement
>>required is 45%. Krait/A15 should get there.
>
>Yeah, because 45% is easy to make up.
>
>Wilco, the "last 10%" is as hard as the first 90% was.
>
>Christ, according to your logic, AMD would have overtaken
>Intel decades ago. Remember back in the 90's when those
>cheap AMD designs were always just 20% slower than Intel,
>so they sold for considerably less?
>
>That was only 20% slower, not 45% slower. So how come AMD
>didn't overtake Intel years ago?
>
>Also, I have news for you: when you try to clock things
>higher, your IPC goes down. Only the bogus benchmarks that
>embedded people still sometimes use (dhrystone)
>
>Finally, it's not true that you can use multiple cores to
>make up for performance problems. Never has been, and
>outside of GPU's, never will be.
>
>Linus
I hope to see the day when Yale Patt's thesis on the systemic design of computing is taken more directly and at face value.
Linus Torvalds (torvalds@linux-foundation.org) on 5/5/12 wrote:
---------------------------
>>
>>According to the above results, the A9 has 31% lower IPC than Bobcat, ie. the improvement
>>required is 45%. Krait/A15 should get there.
>
>Yeah, because 45% is easy to make up.
>
>Wilco, the "last 10%" is as hard as the first 90% was.
>
>Christ, according to your logic, AMD would have overtaken
>Intel decades ago. Remember back in the 90's when those
>cheap AMD designs were always just 20% slower than Intel,
>so they sold for considerably less?
>
>That was only 20% slower, not 45% slower. So how come AMD
>didn't overtake Intel years ago?
>
>Also, I have news for you: when you try to clock things
>higher, your IPC goes down. Only the bogus benchmarks that
>embedded people still sometimes use (dhrystone)
>
>Finally, it's not true that you can use multiple cores to
>make up for performance problems. Never has been, and
>outside of GPU's, never will be.
>
>Linus
