Re: Intel Atom Processor
Marc Olive put forth on 2/17/2010 1:33 AM:
> El Tuesday 16 February 2010 16:52:06 Stan Hoeppner va escriure:
>> As a bonus, due to various architectural reasons I won't delve into, 32bit
>> binaries will usually run slightly faster than the 64 bit cousins
>
> Really? Didn't know.
> 64bit binaries should be faster than a 32bit one... where's the problem?
Instruction word size. On average, depending on individual instruction
encoding, the word length for 64bit binaries can be close to double the word
length of 32bit binaries. What do you think the effect of this is on L1 and L2
instruction caches? That's right, you can fit many more of the 32bit binary
instructions in the caches than 64bit instructions. There are plenty of rename
registers available, and when combined with the extra cache loading, the slight
performance boost allowed by the extra 8 GPRs of 64bit mode is negated and 32bit
code pulls slightly ahead.
The only real advantage of 64bit mode for the vast majority of desktop and
server applications is more efficient addressing of memory beyond 2GB. 32bit
mode has to use PAE which is very inefficient compared to direct 64bit
addressing. For any workstation or server that has 2GB of memory or less, 32bit
OS and apps will be slightly faster than 64bit. We're only talking 1-2% per
core here, basically not noticeable at the application level. You have to run
timed benchies to see a 1-2% difference.
Like I said, 32bit mode will be "slightly" faster than 64bit mode. I didn't say
it would be perceptible. ;)
--
Stan
Reply to: