4 ms·

I'm am very appreciative to Apple for finally closing the endless "The ISA doesn't matter" nonsense. The ISA does matter and it's part of what enabled Apple to

I'm am very appreciative to Apple for finally closing the endless "The ISA doesn't matter" nonsense. The ISA does matter and it's part of what enabled Apple to go so wide. However, there the article is slightly misleading. The ROB isn't where instructions are issued from, that would be the schedulers which usually holds a much smaller set of instruction (16 per scheduler is common). The ROB holds everything that was decoded and not yet retired, thus including instruction that haven't yet been issued to schedulers and more importantly, instruction that have been executed for not retired (eg. might be waiting on older instruction to retire first).

saagarjha 6y ago
Historically, I have yet to see a single case where the ISA has mattered. Every ISA will of course take advantage of its unique features, but in general fast chips come from companies with money and talent, not a particular ISA.
- ridiculous_fish 6y ago
  I think Itanium mattered. Its EPIC instruction set tried to push the scheduling complexity from hardware into the compiler, which proved too difficult. That's part of why it failed.
  - FullyFunctional 6y ago
    Great example. Seeing as I'm working on my Merced box I felt like commenting on this. Intel/HP did many things wrong when they designed EPIC. The absence of dynamic scheduling was supposed to have made it simpler and thus have a faster cycle time. Alas it was so bloated when it finally arrived and horrendously complicated that it clock very slowly (mine @ 733 MHz while contemporaries ran at 1+ GHz). The EPIC designers probably also didn't anticipate the advances in OoO machines that got really good. Trying to statically schedule for unknown load-misses isn't just difficult, it's impossible without code explosion. EPIC got some things right, but design by committee and betting again dynamic scheduling doomed it.

saagarjha 6y ago

Historically, I have yet to see a single case where the ISA has mattered. Every ISA will of course take advantage of its unique features, but in general fast chips come from companies with money and talent, not a particular ISA.

ridiculous_fish 6y ago
I think Itanium mattered. Its EPIC instruction set tried to push the scheduling complexity from hardware into the compiler, which proved too difficult. That's part of why it failed.
- FullyFunctional 6y ago
  Great example. Seeing as I'm working on my Merced box I felt like commenting on this. Intel/HP did many things wrong when they designed EPIC. The absence of dynamic scheduling was supposed to have made it simpler and thus have a faster cycle time. Alas it was so bloated when it finally arrived and horrendously complicated that it clock very slowly (mine @ 733 MHz while contemporaries ran at 1+ GHz). The EPIC designers probably also didn't anticipate the advances in OoO machines that got really good. Trying to statically schedule for unknown load-misses isn't just difficult, it's impossible without code explosion. EPIC got some things right, but design by committee and betting again dynamic scheduling doomed it.