Subject: Speed of 21066 (was: XFree86 on NetBSD/alpha)
To: Dave McGuire <>
From: Joerg Czeranski <>
List: port-alpha
Date: 10/22/1999 09:43:55
Dave McGuire wrote:
> On Thu, 21 Oct 1999, Andrew Gillham wrote:
> >Well, I suppose I'm being a bit harsh, and I haven't tested it in
> >about 1.5 months, so things may be better now. :-)  It seems like a
> >36Mhz SS10 with a CG6 should be slower than a 166Mhz Multia with TGA,
> >but that's definitely not the case. :(
>   No way, man...The Multia (indeed, and every other 21066-based machine I've
> seen) is slower than pissing tar.  I have an article about it somewhere,
> something about memory references int he 21066 being icky-slow or something.
> The '10 will definitely run rings around it.
> [...]

Why do you think that?  For most real life benchmarks (compiling etc.)
my 21066A/233 with 256KB 2nd level cache is twice as fast as our
computing center's SS10 machines (33 MHz).

E.g. the numbers for maxflow (I wrote about that benchmark on this
list, many memory references and a big additional penalty for Alpha:
64 bit int, while the Sparc code uses 32 bit int):
  21066A 233MHz: 267.60s,  SS10 33MHz: 468.05s

Maybe the fastest available SS10 are not much slower than a 21066/166,
but I can't imagine how they could be even faster.  (I don't know the
clock range of the SS10 models, but it certainly doesn't go beyond 66,
or does it?)