Re: Alpha 500 vs Bi Ppro.

Richard Henderson (
Fri, 16 May 1997 13:32:58 -0500 (CDT)

> For exemple we compiled Python 1.4
> on bip (Bi pentium pro) => 26 sec.
> on Alpha => 57 sec
> Ok, maybe the code generation on alpha is longer ?

Quite a bit longer, actually. It is unfortunate, but there are
several algorithms that are used in GCC that blow up exponentially.
One of these involves conditional move which is not used by gcc
on x86. I have a 20-line test example somewhere that pings this
but good -- it takes about 5 hours to compile on my 21164/300. :-(

> python (a little loop program, not very serious abot real
> performance)
> 1) Pic (1 processor only) 1600 pystones
> 2) Alpha 500 under linux 1500 pystones
> 3) Alpha 250 under Digital Unix 1300 pystones

I'd be very surprised if you'd gotten better performance out of this
test on a PC164. This test is almost completely memory bound and the
PC164 board does not have the memory bandwidth of a brand new PPro

> We compiled bogomips (bogo-1.2), and launched 256 in the same time,
> then launched xdvi & ghostscript, and compilations in the same time .

Hmm. I'm thinking that the cache effects of running this many concurrent
proceses put it in the same category as the previous test.

> We didn't test floating point yet, it's not a priority for us. Maybe
> this is a point where Alpha can blast the Ppro ?

The Alpha can blast the PPro in any test that lets it utilize its
larger register set. Poke memory too many times and, with the
exception of very expensive DEC boxes, current Alpha boards will
begin to loose again.

I expect the new 21164PC-based motherboards due out this summer to
rectify that somewhat. The bandwidth on/off chip is about double,
and the new PYXIS i/o support chip can talk to SDRAM.

We'll see.