Date: Wed, 24 Jan 1996 18:45:51 +0100 From: se@zpr.uni-koeln.de (Stefan Esser) To: se@zpr.uni-koeln.de, bem@oce.nl, zawodny@arbd0.larc.nasa.gov, henddjee@cetus.zrz.tu-berlin.de, jbotz@orixa.mtholyoke.edu, <sschmidt@hydra.informatik.uni-ulm.de>, iain@nwpeople.demon.co.uk, freebsd-hardware@freebsd.org, didier@omnix.fr.org, dk+@ua.net, Dmitry Kohmanyuk <dk@dog.farm.org>, s106275@cs.tut.fi (Anssi Saari), "Christof Tebbe" <RZ94-004@wsrz1.wiso.uni-erlangen.de>, duanec@eskimo.com (duane charron), "WanJohn" <WanJohn@email.averydennison.com>, mc100314@macau.ctm.net, phall@umr.edu Subject: AMD 5x86/133 and ASUS SP3G Message-ID: <199601241745.AA08156@Sysiphos> In-Reply-To: phall@umr.edu "AMD 5x86 FAQ:" (Jan 17, 21:23)
next in thread | raw e-mail | index | archive | help
I just bought an AMD 5x86 for my SP3G and wanted to let you
(i.e. anybody interested in that topic, according to my ASUS
mail folder :-) know about my first experiences.
1) The AMD 5x86 works with a write-through primary cache, only.
This seems to be caused by the Saturn chip set of the SP3G
not really supporting write-back primary cache coherency.
(If you get any further information, please let me know!)
2) Jumper setting:
JP30:2&3, JP31:1 to JP32:1 (not a SL-CPU)
JP33:2&3 (2x ==> 4x for AMD 5x86)
JP34:2&3
JP35:1&2
JP36:2&3 (write-through primary cache)
JP40:2&3 (3.45V +- 0.15V)
3) I used both the AWARD BIOS that came with SDMS 3.04.00 and
the latest BIOS awsg0304. The BIOS doesn't seem to make any
difference. Setting JP36 to connect 1&2 (WB primary cache)
makes even DOS fail to load.
4) The performance is lower than I expected.
I'm seeing between 0 and 100% improvement, depending on the
benchmark used. My real world application (recompiling the
full FreeBSD-current source tree) with the old and the new CPU:
22524.98 real 17368.62 user 3597.68 sys (i486DX2/66)
16795.20 real 11689.23 user 2822.52 sys (AMD 5x86)
This is a factor of 1.5 (using 'user' time as the relevant
speed indicator). For reference the times on a Pentium-133:
13449.10 real 8789.22 user 2073.65 sys
That's another improvement by 33%, but it seems the AMD 5x86
is nearer to the Pentium than to a 486 ...
5) Now some byte benchmark results (SP3G, 16MB RAM, NCR SCSI):
BYTE UNIX Benchmarks (Version 3.11)
System -- FreeBSD x14 2.2-CURRENT FreeBSD 2.2-CURRENT #0: Thu Dec 28 21:02:37 MET 1995 se@x14:/sys/compile/NCR486 i386
Start Benchmark Run: Thu Dec 28 23:47:20 MET 1995
2 interactive users.
i486/66 AMD5x86 factor
Dhrystone 2 without register variables 60933.4 74422.6 lps 1.2 (10s, 6 smpl)
Dhrystone 2 using register variables 61878.3 73518.3 lps 1.2 (10s, 6 smpl)
Arithmetic Test (type = arithoh) 127278.7 254943.7 lps 2 (10s, 6 smpl)
Arithmetic Test (type = register) 8656.9 17387.5 lps 2 (10s, 6 smpl)
Arithmetic Test (type = short) 7493.8 15157.3 lps 2 (10s, 6 smpl)
Arithmetic Test (type = int) 8685.3 17383.9 lps 2 (10s, 6 smpl)
Arithmetic Test (type = long) 8684.1 17337.3 lps 2 (10s, 6 smpl)
Arithmetic Test (type = float) 5515.5 11042.3 lps 2 (10s, 6 smpl)
Arithmetic Test (type = double) 5510.2 11033.6 lps 2 (10s, 6 smpl)
System Call Overhead Test 21828.8 27469.2 lps 1.3 (10s, 6 smpl)
Pipe Throughput Test 8522.0 11902.6 lps 1.4 (10s, 6 smpl)
Pipe-based Context Switching Test 3684.7 6212.4 lps 1.7 (10s, 6 smpl)
Process Creation Test 303.2 401.1 lps 1.3 (10s, 6 smpl)
Execl Throughput Test 225.7 331.1 lps 1.5 (9s, 6 smpl)
File Read (10 seconds) 59190.0 70857.0 KBps 1.2 (10s, 6 smpl)
File Write (10 seconds) 6919.0 3800.0 KBps * (10s, 6 smpl)
File Copy (10 seconds) 2439.0 3409.0 KBps * (10s, 6 smpl)
File Read (30 seconds) 60184.0 71891.0 KBps 1.2 (30s, 6 smpl)
File Write (30 seconds) 7088.0 3533.0 KBps * (30s, 6 smpl)
File Copy (30 seconds) 2347.0 3457.0 KBps * (30s, 6 smpl)
C Compiler Test 75.9 88.6 lpm 1.2 (60s, 3 smpl)
Shell scripts (1 concurrent) 112.3 261.0 lpm 2.3 (60s, 3 smpl)
Shell scripts (2 concurrent) 57.0 135.3 lpm 2.4 (60s, 3 smpl)
Shell scripts (4 concurrent) 28.0 70.3 lpm 2.5 (60s, 3 smpl)
Shell scripts (8 concurrent) 14.0 34.7 lpm 2.5 (60s, 3 smpl)
Dc: sqrt(2) to 99 decimal places 6093.1 8567.6 lpm 1.4 (60s, 6 smpl)
Recursion Test--Tower of Hanoi 661.7 668.7 lps 1 (10s, 6 smpl)
INDEX VALUES i486/66
TEST BASELINE RESULT INDEX
Arithmetic Test (type = double) 2541.7 5510.2 2.2
Dhrystone 2 without register variables 22366.3 60933.4 2.7
Execl Throughput Test 16.5 225.7 13.7
File Copy (30 seconds) 179.0 2347.0 13.1
Pipe-based Context Switching Test 1318.5 3684.7 2.8
Shell scripts (8 concurrent) 4.0 14.0 3.5
=========
SUM of 6 items 38.0
AVERAGE 6.3
INDEX VALUES AMD5x86
TEST BASELINE RESULT INDEX
Arithmetic Test (type = double) 2541.7 11033.6 4.3
Dhrystone 2 without register variables 22366.3 74422.6 3.3
Execl Throughput Test 16.5 331.1 20.1
File Copy (30 seconds) 179.0 3457.0 19.3
Pipe-based Context Switching Test 1318.5 6212.4 4.7
Shell scripts (8 concurrent) 4.0 34.7 8.7
=========
SUM of 6 items 60.4
AVERAGE 10.1
(Have a look at http://www.silkroad.com/linux-bm.html for byte bench
results on 486 and Pentium class CPUs under Linux for comparison.)
The arithmetic tests take full advantage of the higher clock rate.
Dhrystone (and worse: Towers of Hanoi) seem memory (or secondary
cache) limited.
The DC test (and the C test) are dominated by process start overhead,
since it takes only 0.007 seconds for DC to find sqrt(2) to 99 digits.
The file copy times cannot be directly compared, since in the last two
weeks there have been a few changes to the file system strategy, which
seem to cause the very different results.
I'm not sure what made the shell scripts performance go up by a factor
of more than 2, but I can reproduce the results.
Regards, STefan
--
Stefan Esser, Zentrum fuer Paralleles Rechnen Tel: +49 221 4706021
Universitaet zu Koeln, Weyertal 80, 50931 Koeln FAX: +49 221 4705160
==============================================================================
http://www.zpr.uni-koeln.de/~se <se@ZPR.Uni-Koeln.DE>
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199601241745.AA08156>