decent performance from G4 Macs?

Mark Hahn hahn at physics.mcmaster.ca
Sun Apr 14 10:43:43 PDT 2002


> On Sat, Apr 13, 2002 at 02:18:39PM -0400, Mark Hahn wrote:
> > is it just that the performance Apple brags about is strictly
> > in-cache, and/or when doing something ah specialized like 
> > single-precision SIMD (altivec/velocity engine)?
> 
> it's the altivec unit that makes G4s at all interesting.  if you
> aren't using the vector unit, yeah, you won't even come close to x86.   

as far as I can tell, the requirement to think highly of G4 is:
	hand-tuned altivec and a tiny working set
which pretty much excludes any general-purpose scientific computing.

> gcc is multi-platform, sure, but it's optimizer for x86 has received a
> lot of attention, while the powerpc optimizer has not. your

I'm not sure that's true: I read the gcc developers list and see
significant efforts from Apple people.  and remember that lots of 
code is not inherently vectorizable, so would never win big on SIMD.

> observation that gcc 3.1 performance is better shows that focus on
> powerpc optimizations has grown, but yeah, it's going to get less

afaikt, 3.1 improvements are from improved infrastructure, nothing
powerpc-specific.

> you are running on mac os x, yes?  is there any chance you could put
> linux on it?  if your application is making a significant number of
> system calls ( file i/o, network traffic... you know, system calls )

no, I'm really only interested in compute-bound performance.

> also bear in mind that G4s run significantly cooler than their x86
> counterparts, so you might still come out ahead on price/performance,

I've heard Apple/Moto's PR on that, too.  but my recent benchmarking
has made me "think different": the G4 appears to be about the same 
performance as current Intel notebook PIII's.  which, of course,
burn about the same power as G4's...

> where price takes into account initial purchase + cost of running the
> cluster.  

we're in the market for 1-200 CPUs.  it's not obvious to me that it 
matters whether the CPU burns 20 or 50W, since we're already got 
30 KW of Alphas in the room ;)

G4e/1000	21	probably "design" power
PIIIulv/700	 8	"design" power
PIIIt/1113	28	"design" power
P4a/2200	55	"design" power
athxp/1800	66	max power

> so there you go. there are lots of reasons why you'll have to actually
> spend a bit of effort to move to a new architecture.  i hope no one on
> this list finds that idea surprising.  

I certainly do.  powerpc support in gcc is not immature, and the cpu
is supposed to be a general-purpose one.  if my observations are true,
then it's the slowest shipping GP machine, and is only viable if you
can afford to structure your program around its SIMD and cache.

regards, mark hahn.




More information about the Beowulf mailing list