riscy
[Top] [All Lists]

Re: Memory bandwidth/board

To: riscy@sunsite.unc.edu
Subject: Re: Memory bandwidth/board
From: tor@tss.no (Tor Arntsen)
Date: Sat, 4 Dec 1993 23:55:53 +0100
On Dec 4, 11:58am, Bill Broadley wrote:
>The motherboard post mentions 84MB/sec read, 104 MB/sec write bandwidth I 
>assume this is for the 50 Mhz motherboard?
>
>So if we go with the 4200 we would be at 40 Mhz or 67.2/83.2 Mb/sec??
>Or would the bandwidth stay the same?
>
>SGI claimes 267 Mb/sec but a similiar memory interface may not be cost 
>effective (let alone designing it.)
>
>I wrote a memory benchmark program (membench at neurocog.lrdc.pitt.edu 
>~ftp/pub) and was wondering what it would report on a SGI mips 4000 system.  
>Any takers out there?
>
>Alpha 133 Mhz gets 80 Mb/sec
>Hp 735 99 Mhz gets 85 Mb/sec
>486dx2 66 Mhz gets 16 Mb/sec
>
>So obviously 67/83 would be very competitive.  I'm still very curious
>if the SGI's would score 267 though.
>
[...]

On an SGI Indy mips 4000PC 50/100MHz with 8K/8K L1 cache (no L2)  I get
36 or 43 MB/sec (not sure to interpret this) if I've done this right.
Bill, I did (slightly edited)
 ./membench 128 2000
 Time      Stride  reads per     % speed     MB per
 per loop  (ints)  second (in K) improvement  sec
  8.60      4       1860.5       21.0       29.1
  6.88      5       1860.5        0.0       36.3
  5.73      6       1861.5       -0.1       43.6
Hmm, I start to believe that the correct number is 29.1 MB/sec.  Whatever
size I use (as long as it's more than 8 (KB)) I get ~29 MB/sec with a
stride of 4 (i.e. 32 bytes).  After that the speed improvement drops to
around zero.

Tor (tor@tss.no)
 
<Prev in Thread] Current Thread [Next in Thread>