From broadley@neurocog.lrdc.pitt.edu Wed Jun 23 08:01:45 1993
Return-Path: <broadley@neurocog.lrdc.pitt.edu>
Received: from gossip.pyramid.com by dutecai.et.tudelft.nl (4.1/1.34JP)
          id AA06046; Wed, 23 Jun 93 08:01:41 +0200
Received: from sword.eng.pyramid.com 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA13738; Tue, 22 Jun 93 23:01:37 -0700
Received: from goss.pyramid.com
	by sword.eng.pyramid.com (5.61/Pyramid_Internal_Configuration)
	id AA06627; Tue, 22 Jun 93 23:01:35 -0700
Received: from neurocog.lrdc.pitt.edu 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA13716; Tue, 22 Jun 93 23:01:33 -0700
Message-Id: <9306230601.AA13716@gossip.pyramid.com>
Received: by neurocog.lrdc.pitt.edu
	(1.37.109.4/16.2) id AA25704; Wed, 23 Jun 93 02:01:16 -0400
From: Bill Broadley <broadley@neurocog.lrdc.pitt.edu>
Subject: Mips 3000 vs 486 (int and fp)
To: riscy@pyramid.com
Date: Wed, 23 Jun 1993 02:01:16 -0500 (EDT)
X-Mailer: ELM [version 2.4 PL21]
Content-Type: text
Content-Length: 4318      
Status: RO
X-Status: 


Okay I'm baised I want cheap mflops at home here's my reasons:

I think the fpu version of the mips 3000 should be used.  While the mips
is an attractive architecture it will have less software available for it.
So it needs an advantage over say just buying a 486 motherboard. 

Including a fpu would:  Prevents re-implementing a floating point math 
library is c-code (which could be WELL spent in other areas), gives the 
board a BIG advantage over 486's (mainly fp), decrease price for people that 
need fpu (could buy in volume, and not have a useless fpu-less mips 3000 
laying around).

Added cost??  Not sure on this one, doesn't the fpu version also include a 
bigger cache?  (this is a BIG win if we don't have a second level cache.)

Theres a bunch of benchmarks at marlin.nosc.mil in pub/aburto, I like them
because normal people compile them with normal compilers as opposed to
"published" benchmarks:

Here's a few samples:  (notice that slower mips 3000's are adjusted for)

Heres a bunch of benchmarks (summary fp on mips 5 times faster, int
performance is similiar for a 486-66 and a mips 3000 at 40 Mhz)

Flops:  A good summmary of FP.  Mips 3000 by a factor of 4.9
	A integrated mips 3000+fpu might be faster then 3000+3010
	
System           OS, Compiler                CPU/FPU    FPU  MFLOPS(1)  REF
                                                       (MHz)               
---------------- ------------------------- ----------- ----- --------- ----
SGI 4D/420       one processor, cc -O1     R3000/R3010  40.0   18.1572    3
Gateway DX2-66   NOTE 041, LINUX 0.99      80486DX2-66  66.7    3.7000   72

Drystone mips:   An int benchmark adjusting for Mhz (33 to 40) mips 1.13 faster
    System                  OS              CPU     (MHz)  V1.1   V2.1  REF
--- ---------------------- ------------ ----------- ----- ------ ------ ---
025 Gateway DX2-66         LINUX 0.99   80486DX2     66.7    0.0   30.9   2
026 DECstation 5000/133    Ultrix 4.3   R3000        33.    31.6   29.0  11

Towers of Hanoi Puzzle Test Results. MIPS 1.70 faster
                                                        CPU   Moves in
System                OS, Compiler               CPU   (MHz)   25 usec ref
--------------------- ----------------------- -------- ------ -------- ---
SGI Indigo            NOTE 012, Irix 4.0.1A   R3000A     33.3   26.999   3
80486/50              NOTE 022                80486DX    50.0   19.056  12

Heapsort:   Mips 1.04 faster
Results as of 05 Mar 1993:                                       HIGH
                                                          CPU    HEAP
System                 OS, Compiler               CPU    (MHz)   MIPS  REF
---------------------- ------------------------ -------- ----- ------- ---
DEC DECstation 5K/240  Ultrix4.2A,cc -DUNIX -O2 R3000     40.0   22.75   9
AMI 80486DX2/66, EISA  NOTE B11, MS DOS 5.0     80486DX2  66.7   21.82  19

Nsieve: mips 1.5 times faster (for low mips)
                                                     CPU    High  Low
System           Compiler                    CPU    (MHz)   MIPS  MIPS  REF
---------------- ------------------------- ------- ------ ------ ----- ----
SGI Iris 4D/35   Irix 4.0.5, cc -DUNIX -O3 R3000     36.0   28.4  22.8   33
Gateway DX2-66   NOTE 011, LINUX 0.99      486DX2    66.7   19.3  16.9   36

Not sure if I believe the below Intel plays some strange games to
get HIGH performance (53 mips for a 486-66???? I never saw more then 30), 
but here it is... IMHO these numbers are unrealistic.  I'm sure SGI
spent less time getting those numbers then Intel did.

Specint:  486-66 1.24 times faster then mips 3000 at 40 Mhz. (adjusted)
    System                      CPU     CPU    GCC   ESP    LI   EQN  SPEC
                                       (MHz)                         int89
--- ----------------------- -------- ------- ----- ----- ----- ----- -----
005 Intel 486 66DX2         80486DX2   66.67  30.5  31.5  48.4  28.8  34.0
011 SGI 4D/320S                R3000   33.00  21.5  23.6  25.5  20.0  22.6

I'm not a hardware guy but I plan to help this project anyway I can.

-- 
Bill					1st>	Broadley@neurocog.lrdc.pitt.edu
Broadley@schneider3.lrdc.pitt.edu <2nd 	3rd> 	             Broadley+@pitt.edu
Linux is great.         Bike to live, live to bike.                      PGP-ok


 
From mackinla@cs.curtin.edu.au Wed Jun 23 09:28:27 1993
Return-Path: <mackinla@cs.curtin.edu.au>
Received: from gossip.pyramid.com by dutecai.et.tudelft.nl (4.1/1.34JP)
          id AA06172; Wed, 23 Jun 93 09:28:20 +0200
Received: from sword.eng.pyramid.com 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA03041; Wed, 23 Jun 93 00:28:17 -0700
Received: from goss.pyramid.com
	by sword.eng.pyramid.com (5.61/Pyramid_Internal_Configuration)
	id AA15275; Wed, 23 Jun 93 00:28:13 -0700
Received: from marsh.cs.curtin.edu.au 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA03037; Wed, 23 Jun 93 00:28:00 -0700
Received: from vincent.cs.curtin.edu.au by marsh.cs.curtin.edu.au with SMTP id AA20445
  (5.67a/IDA-1.4.4 for <riscy@pyramid.com>); Wed, 23 Jun 1993 15:26:04 +0800
Received: by vincent.cs.curtin.edu.au id AA16652
  (5.65c/IDA-1.4.4 for riscy@pyramid.com); Wed, 23 Jun 1993 15:25:40 +0800
From: Patrick Mackinlay <mackinla@cs.curtin.edu.au>
Message-Id: <199306230725.AA16652@vincent.cs.curtin.edu.au>
Subject: Re: Mips 3000 vs 486 (int and fp)
To: riscy@pyramid.com (R3000 PC Mailing List)
Date: Wed, 23 Jun 1993 15:25:40 +0800 (WST)
X-Mailer: ELM [version 2.4 PL20]
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Content-Length: 1035      
Status: RO
X-Status: 


>Added cost??  Not sure on this one, doesn't the fpu version also include a 
>bigger cache?  (this is a BIG win if we don't have a second level cache.)

I believe it does include a larger cache.

[...lots of benchmarks deleted...]

Looking at some old mail from Steve, I see numbers of US$180 for the
40MHz '51 versus US$450 for the 40MHz '81. Obviously, the numbers have
probably changed since then, but the big gap is probably still there.
Seeing as we have a US$500 "cap" on the board, using the '81 would
severely limit the design...

On the general topic of benchmarking 486dx66's versus R3Ks: one thing
to remember is that the R3K is the _old_ MIPS architecture - equivalent
to the 386 in x86 terms. MIPS has most definitely moved onto the R4K
now, which offers >Pentium performance. Also note that the R4K _can_
execute the older 32 bit R3K code...

Pat -- "There's only one thing left to do Mama, I got to ding a ding dang
	my dang a long ling long" (Jesus Built My Hotrod -- Ministry)
GCS d* -p+ c++ l++ m--- s+/- !g w- t- r

 
From andy@resi.waldorf-gmbh.de Wed Jun 23 09:52:55 1993
Return-Path: <andy@resi.waldorf-gmbh.de>
Received: from gossip.pyramid.com by dutecai.et.tudelft.nl (4.1/1.34JP)
          id AA06355; Wed, 23 Jun 93 09:52:51 +0200
Received: from sword.eng.pyramid.com 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA04111; Wed, 23 Jun 93 00:52:48 -0700
Received: from goss.pyramid.com
	by sword.eng.pyramid.com (5.61/Pyramid_Internal_Configuration)
	id AA16839; Wed, 23 Jun 93 00:52:45 -0700
Received: from mail.Germany.EU.net 
	by gossip.pyramid.com (5.61/OSx5.1a Pyramid-Internet-Gateway)
	id AA04107; Wed, 23 Jun 93 00:52:42 -0700
Received: by mail.Germany.EU.net(EUnetD-2.2.6.c) via EUnet
	id IC29073; Wed, 23 Jun 1993 09:49:49 +0200
Received: from resi 
	by wegy.waldorf-gmbh.de with SMTP (5.61/GEN-1.0.7)
	via EUnet for unido
	id AA09303; Wed, 23 Jun 93 09:41:11 +0200
From: Andreas Busse <andy@resi.waldorf-gmbh.de>
Date: Wed, 23 Jun 93 09:39:32 +0200
Message-Id: <9306230739.AA06056@resi.waldorf-gmbh.de>
Received: by resi.waldorf-gmbh.de (5.61/GEN-1.0.7)
	via EUnet for wegy.waldorf-gmbh.de
	id AA06056; Wed, 23 Jun 93 09:39:32 +0200
To: <broadley@neurocog.lrdc.pitt.edu>, <riscy@pyramid.com>
Subject: Re:  Mips 3000 vs 486 (int and fp)
Status: RO
X-Status: 


Hi there !

> Not sure if I believe the below Intel plays some strange games to
> get HIGH performance (53 mips for a 486-66???? I never saw more then 30), 
> but here it is... IMHO these numbers are unrealistic.  I'm sure SGI
> spent less time getting those numbers then Intel did.

I think you're right. I have seen benchmarks (from Intel, of course :-))
where an i486/33 were faster than an R3000... Have they measured nops ?

Anyway, the FPU-version of the R3051/2 called R3081 doesn't have more
cache, so far I know. And I don't think that it's faster than
a R3000/3010 combination with the same clock. The architecture is
exactly the same, only on a single die.

Andy
-------------------------------------------------------------------------------
Waldorf Electronics GmbH, R&D Department
c/o Andreas Busse
Neustrasse 9-12
D-5481 Waldorf
Phone:  +49 (0)2636-80294
Fax:    +49 (0)2636-80188
e-mail: andy@resi.waldorf-gmbh.de
-------------------------------------------------------------------------------

 
