By: Richard Cownie (tich.delete@this.pobox.com), February 5, 2013 9:44 am
Room: Moderated Discussions
someone (someone.delete@this.somewhere.com) on February 5, 2013 9:22 am wrote:
> > In theory, yes. Except that the result of "just recompile" would probably have
> > been something very slow, since the code was mostly about traversing many GB's
> > of complex data structures in thoroughly unpredictable ways, which an in-order
> > static-scheduled cpu would have struggled with.
>
> LOL, as opposed to the OOOE SPARC you were using?
Dude, why do you think I bothered porting to x86 in the first place ? Because
a 300MHz AMD K6 in a $1000 home PC beat the heck out of a SPARC in 1999.
Intel's x86's also ran it great. And a K8 beat the heck
out of later SPARCs. SPARC sucked for this kind of code. I ported onto OoO
x86's as soon as I could get my hands on them. And bought the parts and built the
darn machines myself when necessary (at one point we had twenty boxes with
300MHz Celeron's overclocked to 450MHz, with hefty Peltier coolers).
It seemed likely, based on talking to other people who tried it, that Itanium would have
been just about the same - inadequate - speed as SPARC for that kind of compiler-ish
no-floating-point complex-data-structure-traversal code. And at least SPARC
had compilers and OS that were totally mature and stable, which at that time
was probably not so true for Itanium.
Nevertheless, with Intel as a customer, we nearly had to do an Itanium port for
political reasons, even though it offered no technical advantage - and a lot of
disadvantages - compared to SPARC. The timely arrival of Intel's x86_64 support
saved me from that fate.
> > In theory, yes. Except that the result of "just recompile" would probably have
> > been something very slow, since the code was mostly about traversing many GB's
> > of complex data structures in thoroughly unpredictable ways, which an in-order
> > static-scheduled cpu would have struggled with.
>
> LOL, as opposed to the OOOE SPARC you were using?
Dude, why do you think I bothered porting to x86 in the first place ? Because
a 300MHz AMD K6 in a $1000 home PC beat the heck out of a SPARC in 1999.
Intel's x86's also ran it great. And a K8 beat the heck
out of later SPARCs. SPARC sucked for this kind of code. I ported onto OoO
x86's as soon as I could get my hands on them. And bought the parts and built the
darn machines myself when necessary (at one point we had twenty boxes with
300MHz Celeron's overclocked to 450MHz, with hefty Peltier coolers).
It seemed likely, based on talking to other people who tried it, that Itanium would have
been just about the same - inadequate - speed as SPARC for that kind of compiler-ish
no-floating-point complex-data-structure-traversal code. And at least SPARC
had compilers and OS that were totally mature and stable, which at that time
was probably not so true for Itanium.
Nevertheless, with Intel as a customer, we nearly had to do an Itanium port for
political reasons, even though it offered no technical advantage - and a lot of
disadvantages - compared to SPARC. The timely arrival of Intel's x86_64 support
saved me from that fate.