Pentium Pro Chipset Fault May Reduce I/O Traffic by 90 Percent
Date: Wed, 20 Mar 96 03:27:17 -0800
From: Peter Langston <psl>
Subject: Pentium Pro Chipset Fault May Reduce I/O Traffic by 90 Percent
[There's something awfully familiar about the following interview... -psl]
Forwarded-by: Keith Bostic <firstname.lastname@example.org>
Forwarded-by: Steve Dekorte <email@example.com>
PENTIUM PRO CHIPSET FAULT MAY REDUCE I/O TRAFFIC BY 90 PERCENT 03.01.96
by Alan Beck, managing editor HPCwire
Santa Clara, Calif. -- A defect in Intel's Pentium Pro Orion chipset, known
as the 82450, may limit the amount of I/O bus traffic to as little as 5
megabytes per second, instead of the expected 25 to 55 megabytes per second.
At least 100,000 such chipsets have been shipped since the inception of the
Pentium Pro last year.
Intel alerted motherboard manufacturers to the problem prior to shipment
of the chipsets; however, a non-disclosure agreement between Intel and the
manufacturers prevented the latter from informing end-users. Rectification
of the error was thus left purely to the discretion and resources of the
The data transfer limitation is most likely to manifest itself in high-end
environments, such as servers. According to information obtained from Intel's
Web site: "Certain system configurations involving programmed I/O devices can
reduce the throughput of the entire I/O system, including that of bus master
devices on a PCI bus. When configured as a network server, a mix of
programmed I/O and mastering devices can produce a throughput low enough to
cause the loss of client connections."
Howard High, spokesperson for Intel Corp. in an interview with HPCwire,
noted that all Pentium Pro systems sold to date contain this flaw in the
chipset. Following are selected questions and answers from the interview.
HPCwire: Have all users that received these systems been notified,
definitively and clearly, of the problem?
HIGH: "No. Because it's not a problem. I think we're failing to connect on
communication. If you don't do anything on the work-around, if you just plug
the chipset in, then you have I/O bandwidth limitation that can be 5
megabytes. If you do the work-around, which all the system and motherboard
companies have been notified about, then they can move right up to 5 to 20-
times the 5 megabyte performance. These people are shipping typically into
the workstation space right now, because they haven't introduced their
HPCwire: So everything centers on what you call the "work-around". But
small companies or companies not geared-up for this work may not have the
ability to do it, right?
HIGH: "They would have the ability to do it. Some may choose not to do it.
We've looked at the people who've received Pentium Pro and the 82450
chipsets. There are only a couple that've chosen not to implement the
workaround. All of the rest -- certainly if they've been working with the
majors -- have no issue at all.
"If somebody has a concern; if they say, "Gee, I wonder if I have this
limitation in my system?" we would ask them to call their system provider,
and that way they can tell them exactly what the configuration is in their
system and what the performance capabilities are. We've been polling our
OEMs, and it has not been an issue for them."
HPCwire: Does Intel guarantee this problem will be fixed?
HIGH: "In what way? We keep coming up against this thing where you feel like
it's there, when in probably 98 percent of the cases it's not there -- in the
sense of the workaround being in place. But if somebody has a concern, and
their system provider cannot handle it and make them feel comfortable, then
we ask them to call Intel, and we'll help coordinate and make sure they're
happy with their product.
"I can't say we're going to go back in and rework the chipset, because that's
not how you fix it."
HPCwire: But you just said that the work-around was the only way to fix it.
HIGH: "Right. But depending on how they've implemented the system, there are
a number of different ways you can solve it. It's not: here's one way, the
only way, to do it."
Intel's customer service number is 1-800-628-8686. Information about this
specific problem can be obtained from the World Wide Web at
HPCwire subscribers can obtain further commentary on this issue by ordering
HPCwire article 9087 "OPEN LETTER TO ANDY GROVE RE: THE ORION I/O ISSUE" by
Norris Parker Smith, editor at large.
HPCwire has released all copyright restrictions for this item. Please feel
free to distribute this article to your friends and colleagues. For a free
trial subscription, send e-mail to firstname.lastname@example.org.
© 1996 Peter Langston