Hi,

> Interesting to compare the performance numbers given for the Itanium
> running Glucas v2.8c against my Thunderbird running mprime v21.4 :
>

AFAIK, mprime v21.4 now uses prefetch hints to avoid idle cycles waiting for 
new data. Glucas/Itanium (C-plain code) uses a kind of preload, no prefetch, 
it loads in a loop-cycle what it will need in the next.

This way of preload I think is less efficient than directly prefetch if 
prefetch hints are well tuned. To say the true, it was the first thing I 
tried, but I got difficulties: gcc compiler is still no good for inline 
assembler on Itanium, and Intel compiler simply does not support inline 
assembler (for IA64). So I had to write the preload scheme I told.

But what if FFT data are big enough? prefetch request are some often out of 
L2 memory and the system have to wait many cycles to get the requested data 
near the processor. So, if we've not prefetched the data many cycles before 
we still have to wait some clocks to use.

>  -  At the smallest FFT length, the Itanium is WAY faster.
>

at short FFT length, both mprime/prefetch and Glucas/preload are efficient. 
L2 cache are big enough too. Here, Itanium has the advantage of its superb 
FPU units. It can do any combination of FADD and FMUL up to two ops/clock. On 
the AMD we can only do an FADD and a FMUL in a clock cycle. It is not 
permitted two FMUL.



>       this performance difference decreases until
>
>  -  At FFTs 640K-2048K, the Itanium is a little bit faster
>

Here, perhaps, the mprime/prefetch scheme is better.


>       then the performance difference increases until
>
>  -  At the largest FFT length, the Itanium is noticeably faster
>
>
> Is it memory-bandwidth that lets the Itanium pull ahead at the
> large FFT lengths ?

Now we should need large L2. Some memory request are to main memory. Here the 
L2 size and bandwidth are important. And, just a speculation, prefetch hints 
are less effective.

Guillermo


-- 
Guillermo Ballester Valor
[EMAIL PROTECTED]
Granada (Spain)

_________________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ      -- http://www.tasam.com/~lrwiman/FAQ-mers

Reply via email to