Processor Whispers - About Omitted Features and Own Goals

by Andreas Stiller

Lots of new processors, with 4, 6, 8, 12 ... cores: the Xeon Westmere-EP has arrived, the Opteron 6000 and the Xeon Nehalem-EX are almost there. And, as if there weren’t enough processor versions already, Intel still tops it by offering numerous variations – in some cases with undesirable consequences.

Here in Hanover, scoring own goals is nothing out of the ordinary. The local top division soccer club even scored a near record-breaking three in a single match a short while ago (well, the Belgian Stan Van Den Buys of the Germinal Ekeren club from Antwerp already managed to do that on his own once ...). But also the processor makers are getting better at uncontrollably kicking the ball into the wrong direction.

Intel, for instance, is becoming more and more entangled in the differentiation jungle with their obsessive efforts to artfully bestow different processor families that, ultimately, are all made of the same silicon, with different features.

ECC only in the Afternoon

Of course, people have grown accustomed to the Celeron reaching lower clock rates and having smaller caches as well as slower memory than, for example, a Core i7. However, Intel also insists on differentiating those features that the average customer practically doesn’t (want) to know about: hardware virtualisation (missing from many Atoms), SSE4.2 and AES instruction set extension (missing from the Pentium G, the new Celeron G) and ECC memory protection (missing from the Core i7). This distinction between the smallest details does not only confuse PC owners, who actually might want to use one of these features, which sometimes are indeed supported by the processor but blocked by the motherboard BIOS, but also confuses software developers.

Now Intel has gone one step further and – without any compelling technical reason – is now making certain CPU features dependent on the chipset.

According to Intel’s information website ark.intel.com, the low-budget processor Celeron G1101, for instance, supports ECC while the much more expansive Core i7-860 does not. The reason: the Celeron is also intended as a low-budget option for the application with LGA1156 server boards while clients are supposed to buy the Xeon X3460 instead of the cheaper, but almost identical – apart from the ECC support – Core i7-860. However, the Celeron will not support ECC when used with a normal desktop PC motherboard. Although it has the memory controller, ECC memory protection is meant to work only in connection with motherboards that have the 3400 or 33420 chipset and not with those that, for example, have the H55 chipset. Again, there’s no technical reason for this: the chipsets 3400, 3420 and H55 are practically identical and, actually, are only southbridges that don’t have anything to do with the memory anyway.

Maybe soon ECC will only work in the afternoon between two and four o’clock for some configurations, the hyper-threading will be enabled on Sundays and the extended virtualisation capabilities will be available at night – but only by moonshine.

Opteron 6000 alias Magny-Cours: finally a large non-square processor again – like the Pentium Pro once was Microsoft, for one, has apparently grown tired of the continuous inquiries from Windows 7 users concerning the hardware requirements for the “XP Mode” and has now ceased to require processors to support virtualisation commands. A vintage own goal for Intel as, now, there is one less reason for buying expansive (Intel) processors. And it’s likely that other software developers will also refrain from using new CPU features as long as Intel continues to create so much confusion about this matter.

Like with some other processor families before, some special versions of the new Xeon processors with the Westmere core will be marketed as embedded processors, which means different delivery and support conditions and, partly, other thermal and electric specifications. Of the “normal” versions, the most outstanding is the strongly degraded Xeon L5609 at $440: without turbo boost, without hyper threading, without thermal monitor and with a slower QPI (4.8 GT/s), but also with AES, TXT and VT ...

48-Core Contests

Meanwhile, Intel and AMD have taken rather different approaches concerning the topic “48 cores and what you can do with them”. AMD had advertised – limited to the USA and Canada – a Tyan system with four Opteron 6174s and a total value of $8189 as reward for the best contribution (in the form of an essay or blog with up to 500 words or a video of up to 3 minutes) that presents sensible options for the application of such an amount of cores. The 12-core Opteron 6000 (Magny-Cours), which – in a set of four –will allow for servers with 48 cores, is supposed to officially roll out on the 29th of March. Already, various prices have leaked from online shops, which – ranging from $293 to $1491 – have been arranged in a rather aggressive way. Their purpose is to give Intel’s 8-core Nehalem-EX, which coincidentally will be launched the day after, a hard time.

Those who can convincingly explain what they plan to do with it have a chance of receiving a test board with the experimental 48-core chip SCC from Intel (here in the experimental platform Copper Ridge). With its experimental processor SCC (single-chip cloud computer), codenamed Rock Creek, which has mostly been developed in Braunschweig, Germany, Intel has packed 48 cores on a single chip. And there’s a kind of contest involved, too. It’s possible to apply in order to receive a SCC system for highly meaningful research purposes. This offer should be especially interesting for Universities and industrial research divisions for experimenting with this flexible mixture of cluster and SMP system, in which the cores are linked on-board through a very fast, configurable mesh. However, time is of the essence: all “SCC Research Proposal Applications” have to arrive at Intel by the 15th of April.

(djwm)