Subject: Re: Beautiful Sparc, take 2
To: doomwarrior <doomwarriorx@gmail.com>
From: Timo Schoeler <timo.schoeler@riscworks.net>
List: port-sparc
Date: 05/30/2006 17:44:33
>> HyperSPARC *needs* high clock frequencies, as can be seen here, to
>> perform well.
>>
> hmm i got a diffrent result. All tested with a SparcStation 20 with
> 512MB RAM and Seagate ST32550WC.
> I always run each test 3 times for each processor and throw away the
> worst and the best result.
>
> time bzip2 --best gcc-4.0.tar
> 1x Ross HyperSparc RT620B \w RT626 125 1398.692sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
> 1x Sun SuperSparc II \w SuperCache 75 1798.430sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
> 1x Sun SuperSparc \w SuperCache 60 2278.572sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
> 1x Ross HyperSparc RT620A \w RT625 55 2464.82sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
>
> i think the RT620B start to suffer from disk i/o-limitations, because a
> RT620D with 180Mhz isn't really faster.
>
> time oggenc -m 160 -M160 MikeOldfield-TubularBellsPart1.wav
> 1x Ross HyperSparc RT620B \w RT626 125 8692.360sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
> 1x Sun SuperSparc II \w SuperCache 75 16402.050sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
> 1x Ross HyperSparc RT620A \w RT625 55 18890.194sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
> 1x Sun SuperSparc \w SuperCache 60 19568.501sec (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
>
> I think the values speak on there own. The RT620B is twice as fast as
> the SuperSparcII.
>
> ./Run dhry2reg (UnixBench 4.1.0)
> 1x Ross HyperSparc RT620B \w RT626 125 210999.300 (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
> 1x Sun SuperSparc II \w SuperCache 75 121819.300 (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
> 1x Sun SuperSparc \w SuperCache 60 96130.500 (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=supersparc)
> 1x Ross HyperSparc RT620A \w RT625 55 93506.100 (gcc
> 3.3.3nb3/NetBSD 3.0/-O3 -mcpu=hypersparc)
>
> With the huge L2 Cache on all processors, this test should be completly
> inside the L2-Cache.
> The INT-performance is nearly the same.
thanks for the numbers. i saved 'em in my archive :)
> don't trust benchmarks you don't manipulate on your own.
if churchill had known computing, he'd given us the appropriate statement ;)
> Also have a
> look at the internal differences between
> HyperSparc and SuperSparc.
'The level-2 cache on HyperSPARC modules has a lower latency
(zero-CPU-bus-wait-state load/store access for cache hits) than that
that found on cached SuperSPARC modules.' etc.
but you know this site (mbus.sunhelp.org), as i just saw =8-)
> The HyperSparc has _2_ FPUs, the SuperSparc
> has only 1. The SuperSparc is a bit faster then
> a Pentium P54C, the HyperSparc is faster then a Pentium P54C and at the
> same level as a Pentium P55C (MMX)
hm, the P55C has it's (tiny) performance gain from doubled L1 caches,
IIRC, and maybe another organization of associativity of it's cachelines
(just a thought).