The Windows Observer--AMD Stalled by a Bug in Barcelona Opterons

Newsletters	Subscriptions	Forums	Safari	Store	Career	Media Kit	About Us	Contact	Search	Home

Volume 4, Number 45 -- December 12, 2007

AMD Stalled by a Bug in Barcelona Opterons

Published: December 12, 2007

by Timothy Prickett Morgan

We have all done it before. Rushed through our homework and made a math error that messes up a complex calculation and therefore gives us the wrong answer. And so it is with Advanced Micro Devices and its "Barcelona" Opteron processors, which have a bug. Or more precisely, an erratum, in chip-maker speak. And that bug has required some BIOS workarounds in Barcelona systems and caused AMD to pull back the reins on Barcelona until a next stepping of the chip is out in the first quarter of 2008.

Barcelona is, of course, the much-anticipated quad-core implementation of the Rev F Opteron design, complete with extra goodies to make main and cache memories and server virtualization perform better. Barcelona was expected to come to market in June or July of this year, and did not come to market until September. And then it came out at clock speeds that were lower than many had hoped and insufficient to keep a lot of competitive water between the Barcelonas and Intel's quasi quad-core Xeons, code-named "Clovertown" and "Harpertown" in two-socket machines and "Tigerton" in machines with four or more CPU sockets. Intel is already ramping up its quad-core chips in 45 nanometer processes with the new "Penryn" designs (Harpertown is the first Xeon based on this design), and AMD is struggling to get its 65 nanometer quad-core "Budapest" (for two-socket and single-socket machines) and Barcelona processors ramped up.

AMD had been hoping for the Barcelona and Budapest chips to give its business a kick, but the delays in getting these processors out the door have hammered its stock price on Wall Street and has driven the ouster of some top people at the company, including long-time chief of marketing, Henru Richard, just two weeks before the Barcelona launch on September 10. In the wake of the report of the bug in the Barcelona chip, AMD's stock price sank to a 52-week low of $8.79 per share, giving it a market capitalization of around $5 billion. A year ago, AMD's shares were hitting a 52-week high of $23 a pop even with Intel coming back to life in the market for processors for workstations and servers, giving it a market cap of $12.5 billion. AMD's market cap is now lower than the price it paid to acquire graphics chip maker ATI Technologies in July 2006. AMD's fortunes have been rising since the middle of 2004, when Intel's Xeon processors were overheated and underpowered compared to single-core and then dual-core Opteron alternatives; AMD's market cap peaked near $42 a share in early 2006, and has been sliding since then in direct proportion to AMD's own slippages with Barcelona and Intel's own success with the "Woodcrest" and now "Penryn" Xeon cores and their power-conserving Core microarchitecture.

While 2003, 2004, and 2005 were a lot of fun for AMD, as it got endorsements from all of the major server vendors and played into the IT market's increasing focus on performance per watt as much as price/performance, 2006 and 2007 have not been so much fun with a resurgent Intel and product delays and now a bug. All chips have erratum here and there, and often require fixes in operating system microcode or BIOS software to get around them. It is the nature of such complex pieces of technology for a bug to escape detection or for chips to illicit weird behavior under conditions that cannot be predicted once real software with real workloads starts running on the chip. While Barcelona's volumes were never going to be high during the ramp this year, because they are server processors with lots of features for virtualization, they command relatively high prices (compared to desktop and laptop chips, at least), which could have helped AMD's bottom line considerably in 2007.

According to a statement put out by AMD last week, the company is only shipping Barcelona chips to customers who can live with a BIOS fix or have other software workarounds to cope with the bug. Some high-performance computing labs have taken Barcelona shipments, but it is unclear what exact workloads AMD is talking about. The bug has to do with the transition lookaside buffer (TLB) in the Barcelona and the quad-core Phenom Athlon-style desktop chips, and specifically deals with data moving in and out of the L3 and L2 caches that can cause a system crash. (You can read about it in this posting on the X86-64.org site.) AMD has released a kernel patch for Linux to cope with the problem for those who do not want to take the performance hit by using the BIOS workarounds. And the warning for the use of the GPL code that patches Linux kernels is enough to make everyone not put it into production:

Due to the very invasive nature of this patch and the very small number of affected customers (you know it if you have an affected part), we do not recommend the use of this patch on a regular Linux system. This patch is NOT intended for mainline acceptance or inclusion with a Linux distribution! The patch has only received minimal functional testing. Every user must evaluate it prior to production use to make sure it meets the necessary quality standards. Like all GPL software, this patch comes with absolutely no warranty.

Waiting until the first quarter of 2008 sounds like a pretty good idea if you are trying to get the most performance on a Linux-based Barcelona box.

The statement put out by AMD says that the erratum concerning the TLB cache in the Barcelona and Phenom processors was known before the November 19 ship date of the Phenom chips and that a BIOS fix that came out prior to the Phenom launch fixed the issue. "We are experiencing strong Phenom demand and are shipping parts to the channel, system builders, and OEM customers." AMD added in the statement that in addition to "strong demand" for quad-core Opterons, it was seeing strong demand for dual-core Opteron chips (presumably only Rev F variants), and that IDC market share statistics show AMD-based servers gaining market share in the United States and worldwide in the third quarter.