| I found this on the internet, FWIW.
-- Russ
Re: Cache and performance (was Re: Alpha Purchase?)
----------------------------------------------------------------------------
From [email protected] (Robert Harley)
Organization I.N.R.I.A Rocquencourt
Date 25 Apr 1997 18:16:16 GMT
Newsgroups comp.sys.dec,comp.benchmarks
Message-ID <[email protected]>
----------------------------------------------------------------------------
Niclas Mattsson <[email protected]> writes:
>>True, true. However this will become moot when DEC semiconductor
>>releases their new motherboard with SDRAM support, making cache far
>>less of an issue [...]
>Will someone expand on this? I'm considering the new Personal
>Workstation 600a [...] Please deconfuse me!
Our group here at INRIA just got a PWS 500a and we're running Linux on it.
It's fucking awesome!
Here are are some numbers showing the speed (iters/time) on John Tromp's
Fhourstone benchmark, which does fairly random accesses in a large
array. I ran it on an AlphaStation with 8MB of L3 cache and on the
PWS with 2MB. Both have 500MHz 21164a chips. For small sizes the two
are very similar. Here you can see how the AS wins for sizes close to
it's cache size but the PWS overtakes it for big problems.
AS 500 PWS 500a
5MB 618.0 450.2
10MB 542.2 439.0
15MB 486.0 440.0
30MB 432.2 442.9
60MB 407.6 430.3
120MB 391.2 417.1
240MB ----- 401.1
I can run this test on an Aspen clone with 1MB cache too if people are
interested.
The following Stream numbers also show the PWS's advantage (for large
problems with little locality). For a non-cached machine, add 30 or 40 MB/s:
Copy: 269, Scale: 267, Add: 289, Triad: 312
Bye,
Rob.
.-. [email protected] .-.
/ \ .-. .-. / \
/ \ / \ .-. _ .-. / \ / \
/ \ / \ / \ / \ / \ / \ / \
/ \ / \ / `-' `-' \ / \ / \
\ / `-' `-' \ /
`-' Linux - 500MHz Alpha - 256MB SDRAM `-'
|