Re:A Shocking Truth About CSS
by btilly on Tuesday March 09, 2010 @11:20AM
(#1178597)
Billions of cycles per second. But executing an instruction frequently takes 20 cycles or so. That reduces instructions to hundreds of millions.
[ Parent
| on HN ]
The core i7 can issue 4 instructions, and execute 6 micro-ops per cycle. As long as your code is not a single huge dependency chain, you are mostly reading data out of L1 cache, and you don't use complex instructions like division, you can expect to get at least 2 instructions per clock. I have profiled real code (crypto stuff) that got a tad bit over 3 instructions per cycle. That's 10 billion instructions a second on the processor it ran on.
Most instructions have not taken over 2 cycles on x86 since P4. This is a mesofact, update your worldview. :)
Except that pipelining, multiple cores, and hyperthreading then bring it back up. So, let's not get overly pedantic here.
Well, it doesn't happen in real life, but an Intel i7 is capable of issuing 4 instructions per clock cycle (sometimes 5, with a little microcode cheating). A 3.2 GHz quad-core could issue 51.2 billion instructions per second.
Re:A Shocking Truth About CSS
by jacoblyles on Tuesday March 09, 2010 @11:50AM
(#1178707)
Presumably any modern processor would use some form of pipelining, which I think would take instructions back up into the billions. (The pentium 4, if I recall correctly, had a 20+ stage pipeline)
Re:A Shocking Truth About CSS
by bruceboughton on Tuesday March 09, 2010 @12:17PM
(#1178805)
From CSS to processor design in just 5 comments! Only on HN ;)
Long pipelines cause issues when branches get mis-predicted. The half processed instructions have to be thrown away.
What you want is a CPU with a short pipeline, a fast clock rate and if possible instruction reordering and multiple dispatch.
Large caches and good branch prediction help a lot too.