> earthdog wrote:
> 
>>Oops.  Missed that, didn't I?  But that would explain some of the
>>problem?
> 
> 
> Using different compiles would indeed invalidate the results (unless
> both compiles were done by the same compiler, and optimized for each
> CPU, which was not the case here).
> 

Depends on what you want to test.
For the test they show, if they wanted to see the relative speed
of the processor running old-code application, using the same binary 
distribution for all would have been mandatory, which should have
rules out the Itanium2 measurement (which I do not trust either, yet).
If they wanted to see the relative speed of the processor doing 
something serious and real (and ray-tracing IS a real thing, not
like the classical benchmarks doing loop in the CPU cache), they might 
have include the Itanium2 results, but to be fair they should have done
a custom compile (if available) of each system and application.


> I suspect that an Opteron-optimized compile would do considerably
> better.

I hope so, but it may take more time for it than for the Itanium2: Intel 
make the optimising compiler for that one (but will it be available
at the same price as the Gnu/Linux for everybody, I really doubt that).
I believe an opteron-compile would perform better, and it might
even be 'interesting' to rewrite Povray as multithreaded rendering 
engine (I know the FAQ about that, just dreaming, with a 3 or 4 Opterons 
system, just to keep the memory latency low).
I also believe the Itanium2-version will be out-of-reach for the
mundane like me, and the performance with the Gnu suite are still 
unknown (better or worst than the Opteron, who know ?).

Post a reply to this message