[svlug] [Off-topic] Do NOT buy an Intel i820/840 mainboard/system that uses SDRAM ...

Bryan J. Smith b.j.smith at ieee.org
Fri May 19 22:08:43 PDT 2000


[Off-topic]  Do NOT buy an Intel i820/840 mainboard/system that uses
SDRAM ...

[ I've commented on this on various hardware lists, etc...  And it's
no longer just a performance issue, but a MASSIVE DATA INTEGRITY
ISSUE NOW!  And it's not just the mainboards you have to look out
for, but complete systems from name brand vendors.  I JUST WANT
EVERYONE TO KNOW, because the media is barely covering this.  I am
buy a new server soon and I keep running into tier-1 OEMs and
vendors who sell systems using these defective SDRAM-based i820/840
mainboards.  They are still selling them because Intel has made this
a "voluntary" recall (NOT mandatory, for obvious cost reasons -- ***
SEE STRIKE #3 BELOW *** ]

The Hitter:  Intel 82805AA "MTH" Chip

The Intel i820/840 chipsets are designed for Rambus [Direct] Rambus
DRAM ([D]RDRAM).  To accomodate the "transition" from SDRAM to
[D]RDRAM, Intel designed a "memory translator hub" (MTH), the Intel
82805AA chip, for use on mainboards to convert the SDRAM to RDRAM
signaling so lower-cost SDRAM could be used.  But this was not only
for using lower-costing SDRAM, but because RDRAM is currently only
available in RIMM modules no greater than 256MB, whereas 1GB SDRAM
DIMMS are now available (and 512MB SDRAM DIMMs are quite
commodity).  As such, i820/840 systems with the MTH and SDRAM DIMM
slots are common for servers where a high memory capacity is needed.

Unfortunately for consumers, Intel has swung and missed a 3rd time
and struck out.  This last being a complete and UTTER AND
*UN*RECOVERABLE DESIGN FAILURE for shipping boards that has turned
into a unnoticed, but *** MASSIVE RECALL ***.  But the problem does
NOT STOP THERE.

[ Note:  Strike 1 & 2 are just for historical reference.  Skip to
Strike 3 for more on the Intel recall. ]

Strike #1:  RDRAM Signaling Failure

Admist mainboard production by Intel, major Taiwanese and other
manufacturers, Intel discovered that it's i820 and i840 mainboards
could NOT handle the signaling of 3 RIMM slots.  The signaling
became weak and data losses became common.  Even though RIMM slots
require a "dummy" terminator in empty slots to help keep EMF
sensitivity at a minimum, 3 slots were too subjective to external
noise.  As such, produced mainboards were scratched and redesigned
to support only 2 slots, severely limited memory capacity.

Strike #2:  SDRAM Performance c/o the MTH

Even though the native RDRAM i820/840 mainboards were delayed,
i820/840 mainboards with SDRAM slots and MTH chip(s) were available
at the same time as the RDRAM production halt.  Within days, major
PC enthusiast sites like Tom's Hardware, Anandtech and even major
engineering and trade magazines were reporting extremely pitiful
memory performance.  Most benchmarks not only placed i820/840 + MTH
performance below the 2-year old PC100 i440BX chipset, but even the
3-year old PC66 i440LX chipset.  It wasn't long before most everyone
was recommending VIA's PC133 mainboards over Intel's i820/840
solutions.

Strike #3:  MTH Proven Unreliable, Intel Recalls SDRAM i820/840
Boards

Today, several million i820/840 + MTH SDRAM mainboards have made
their way into OEM and end-user products, with most in use now.  Two
weeks ago, Intel made it know that ALL MAINBOARDS WITH THE I82805AA
CHIP ARE *UN*RELIABLE.  Intel has NO FIX *NOR* ARE THEY GOING TO FIX
THE PROBLEM.  Intel admits that it decided to release the MTH chip
fully knowing the rate of failure and critical memory faults were
extremely high, but no redesign would be possible since the use
something like the MTH (to translate such high speed signals from
one to another) is really the root cause.  The production of the
i82805AA will cease (when though?!) and i820/840 SDRAM mainboards
will no longer be available in the near future.

So what's the fix again?
   ** THERE IS NONE **

So what's Intel going to do about it?
   Intel has a "voluntary" recall on these boards.  Intel, at it's
own expense, will replace the SDRAM i820/840+MTH mainboard with
SDRAM memory with an RDRAM i820/840 mainboard with RDRAM.  Given the
cost of RDRAM, this *IS* going to cost Intel a crapload (assuming
everyone takes them up on it).
   *BUT* a problem arises on i820 SDRAM mainboards which can have
upto 4GB of SDRAM, and i840 SDRAM mainboards with upto 8GB of
SDRAM.  Equivalent RDRAM memory and slots on each can only address
upto 512MB and 1GB currently, respectively, and only 1 and 2GB with
the availability of new 512MB RIMMs.  I have not heard yet what
Intel is doing about that.

** SO WHY SHOULD I CARE? **
   Because the recall is "voluntary", THERE ARE STILL *MAJOR OEMS*
SELLING I820/840 MAINBOARDS WITH SDRAM!  As such, *BEWARE* when you
buy servers and high-powered x86 workstations from various vendors
like Compaq, IBM, etc...  Check to make sure the board is either
powered by RDRAM, or an SDRAM-powered mainboard with a non-i820/820
chipset like those from VIA, ServerWorks and others.  Likewise,
beware of the increasingly low cost i820 SDRAM mainboard -- not only
because of performance reasons, but because of the data corruption
issue.

And to finish off the story ...

The Strikeout:  Industry Embraces DDR DRAM
   Intel is currently trying to get RDRAM production up, and prices
lowered.  Unfortunately, even with mass production, RDRAM is still
costly.  As such, major memory manufacturers, both foreign and
domestic, are pushing forward with a more evolutionary SDRAM
approach, double-data-rate (DDR), already proven by video chip
designers as better than RDRAM for less price.

-- TheBS

P.S.  I work with a lot of former Real3D guys at Theseus, some of
which were hired by nVidia and ATI.  Intel pushed 800MHz DRDRAM on
them and they still chucked it for PC200 DDR SDRAM instead.  Not
surprisingly, even Intel continues to use [DDR] SDRAM on its i810
(with the integrated i752 video logic) and it's forthcoming i815
(with integrated i752 successor with an optional AGPx4 slot).  Don't
get me wrong, serial memory *IS* the future of I/O, *BUT* RDRAM
ended up *NOT* being the "stepping stone" Intel wanted it to be.

-- 
 Bryan J. Smith    mailto:b.j.smith at ieee.org,thebs at theseus.com
     Disclaimer:  http://www.SmithConcepts.com/legal.html
**************************************************************
"[the Digital Millennium Copyright Act (DMCA)] does something
 no copyright law has ever done -- it extinguishes fair use"
            -- Martin Garbus, respected First Amendment Lawyer






More information about the svlug mailing list