AMD loses largest Asia supercomputer deal

WE'VE ALL talked a lot about Barcelona delays, under-deliveries and all the plagues including the infamous "TLB errata".

Try to find Barcelona-based servers around your corporate neighbourhood - you may as well look for a needle in a haystack. Even with these mounting headaches, AMD still managed to keep substantial presence, if not lead above Intel, in the top of the top, the large scale supercomputing arena.

When you look, for instance, at the more recent US government projects, you'll see that a proportion of AMD-based systems is still high, not just because of the Hypertransport and memory bandwidth scaling with each added CPU. The previous Opteron generation had an advantage over Intel platform in 64-bit code when using large page sizes - like in Global Linpack benchmarks, for instance. The Nehalem generation might solve this problem, though.

Also, according to our best friends amongst that user base, there are some known issues in the SSE3+ instructions, not adopted by AMD that cause system crashes with large page files - the US Gov't has a fix, but I'm not so sure of Chinese and Russian supercomputer users relying on the same chip having detected or solved that problem, for that matter. In a few cases, when using a given HPC code on both Intel and AMD 64-bit machines, the AMD64 code had to be tweaked to run on EM64T, even though, in say generic Windoze or Linux, there are usually no problems.

Finally, when, at the end of 2008, you look at the top of the Top 500 list during the next SuperComputing conference, you may see several near-petaflop-class AMD-based US government systems appearing there - names like "Roadrunner", "Baker" and "Ranger" should mean something to those in the know.

However, the accumulated delays and performance target misfirings did affect the confidence of the rest of the world, it seems. We talked recently about AMD's Asian supercomputing wins and their delayed fulfilment. One of those was a large, 150+ TFLOPs central Korean research supercomputer at their famous KISTI institute nestled among the hot springs.

The original AMD-based spec that Sun Microsystems won the deal with some 9 months ago, revolved around many 2.5+ GHz Barcelona quad core CPUs. As you know, a 2.5 GHz Barcelona would, with 10 GFLOPs peak per core, provide 40 GFLOPs peak per chip, or some 4,000 chips (over 500 eight-socket systems) to come close to the performance goal. The above mentioned HPC performance benefits with total bandwidth, SMP scaling and large pages and so on would have a factor in deciding, too - especially since all that happened before Intel's 45 nm stuff was out there.

This would have been the largest supercomputer at the time in Asia, and a major statement of confidence in AMD platform, keep in mind.

We all know what happened next, both with the performance and the deliveries. And, the inevitable followed - the last I heard recently was Sun staff were busily replacing AMD with new 45 nm Intel specs, and the Koreans will, now, have Intel "eight-brain chips" - to use BBC Teletubby-IQ wonderspeak, "thinking" inside their, and probably Asian, largest supercomputer in these few months.

To me, this loss, or should I say a fateful AMD-to-Intel "switch", is more than just a few thousand of chips going the Intel way rather than AMD. The perception that AMD just went overboard with the CPU screwups, and the huge site switch of this proportion is bound to have a snowball effect among many high-teraflop and near-petaflop projects going in North Asia, the world's fastest growing and already second biggest HPC market. So, unless Barcelona B3 changes this suddenly, this is going to be the first switch among many.

Also, this win may embolden Intel to fix the remaining performance drawbacks they had vs AMD in the supercomputing arena. Nehalem or its 32 nm successor, Westmere, may have that done, just like the interconnect and integrated memory were solved. Worst come to worst, just resurrect the EV9 Alpha as QuickPath-based "computation accelerator" for those Nehalems, and voila - you got perfect flat memory model, superb code performance and all the bandwidth in the world, coupled with a proven X86 front end.

Once (not "if") these are solved, what is there to stop Intel from aggressively attacking this last niche of AMD's domination? Yeah, there isn't much real hardware profit in there, but the exposure, the "mine is bigger/faster than yours" statement of strength, and, of course, the government user influence worldwide, is, to paraphrase MasterCard ads, priceless.

Not to forget, South Korea was one of the sites of investigation of alleged Intel malpractices, and, in this situation, AMD getting a deal and misdelivering it, will not help their case ultimately.

Actually, Sun could have stuck with AMD till the end, and possibly kept the deal for the DAAMIT gang, if AMD listened to that same Sun and offered a Barcelona MP version with all four originally planned HT links at HT3 speed, rather than dumbing it down to socket-compatible three HT1 links as "other" (read: more Intel-centric) Opteron vendors requested. The added bandwidth and scaling efficiency in those 8-socket Galaxy boxen may have just convinced the users to wait a little more.

This way, by the time "Montreal" and such CPUs are supposed to come out from AMD sometime in 2009, Intel will already have 4-link QPI on Beckton 8-core Nehalems ready - and, the advantage they could have had for a year, will be nowhere...

As for Sun Microsystems, they are investing big time in supercomputing this year, including hiring tons of people and designing more HPC-centric systems. The problem is, many more of those may be Intel-based this time. µ