[X4U] Re: New Pentium Macs (Altivec)

Vard Nelson vnelson at detectgeo.com
Tue Jan 24 10:57:55 PST 2006


>>  > . . .PowerPC could no longer
>>  > compete, regardless of broken promises from Motorola and IBM.
>>  > Things like the Velocity Engine (i.e. AltiVec) were stopgaps.
>>
>>We do serious number-crunching on the G5s (many jobs run for days).  The
>>Altivec engine has made certain kinds of applications SCREAM such that a
>>single G5 can cometimes compete with a cluster of dozens of Winboxes, and
>>"dozens" is an understatement.  Of course, that has required
>>Altivec-specific programming.  We are NOT looking forward to massive
>>re-writes of our code.
>>
>>Oh, well, the price of wanting to be bleeding edge.

>. . .I've been wondering about applications optimized to use Altivec. Does 
>the Altivec silicon do anything more than accelerate the floating point 
>performance?
>
> The SPECmark performance for FP that Apple is reporting is much better 
> than the Altivec-equipped G5. So if this line of thinking is correct, will 
> it be necessary for your team need to write a lot of low-level code for 
> the Intel Core Duo? or just use a top-shelf optimizing compiler to take 
> advantage of the chip's basic architecture?


Don't know.  We write mostly in Fortran, using the Absoft compiler.  Some 
code we have written is low-level, some is optimized by the compiler.  The 
Altivec engine requires that the number of elements in arrays be divisible 
by 4 (or 8 sometimes??), so there has been a lot of detail work preparing 
the code for Altivec use.  Any changes might require those sorts of 
re-writes.  The stomach churns.  Of course, any quantum-leap performance 
increases will be most appreciated!

Vard Nelson
Detect Geophysical 



More information about the X4U mailing list