The folks at AMD have Valentine's Day gift for the server market today in the form of five new, faster models in the Opteron 6100 series. All of these processors are designed to drop into the double-wide G34 socket with quad memory channels, and they're essentially new top-end parts for each power envelope addressed by the Opteron 6100 series. The table below shows the key specifications for the new CPUs, along with pricing.
|Opteron 6132 HE||8||2.2 GHz||65 W||$591|
|Opteron 6140||8||2.6 GHz||80 W||$989|
|Opteron 6166 HE||12||1.8 GHz||65 W||$873|
|Opteron 6176||12||2.3 GHz||80 W||$1265|
|Opteron 6180 SE||12||2.5 GHz||105 W||$1514|
The new models essentially supplant existing products at the top of each power band while offering slightly higher clock frequencies. The 105W Opteron 6180 SE, for instance, is 200MHz faster than the prior top bin, the Opteron 6176 SE. Others, including the Opterons 6166 HE and 6176, are just 100MHz faster than the previous peak speeds at their power levels. The older models should to drop in price accordingly across the 6100-series lineup.
AMD Product Marketing Engineer Michael Detwiler tells us these new Opterons should be available immediately through component distribution channels. Large server vendors have purportedly been shipping systems equipped with the new speed grades "for the last few weeks," so complete servers based on them ought to be imminently available, as well.
AMD's Opteron business has been facing incredibly tough competition from Intel's 32-nm Westmere Xeons for most of the last year, but AMD continues to pursue a careful strategy of product positioning and marketing. The quad-channel G34 socket gives AMD's server chips a technology advantange that its desktop and mobile parts lack, as Detwiler underscored by citing a roughly 30% performance advantage in SPECint_rate2006 and SPECfp_rate2006 for the Opteron 6180 SE versus the Xeon X5680, Intel's fastest six-core Westmere Xeon. (As of now, we don't see the 6180 SE results published at spec.org. AMD says they were submitted on February 3, 2011.) By offering more cores, memory channels, and even chips at the same price levels as the six-core Xeons, AMD has been able to ensure its products retain some attractiveness for certain workloads and applications. Small victories like those are probably all that AMD can muster until the arrival of its Bulldozer architecture.
Development on the server version of Bulldozer continues to proceed more or less on schedule, though, according to Detwiler. The chip taped out in Q2 of 2010, first sampled to parters in Q3 of last year, and is slated to go into production in Q2 of this year. Bulldozer server processors should be introduced and reach "widespread availability" in the third quarter of 2011. (Although the client and sever versions of the chip more than likely share the same silicon, Detwiler was careful to point out that the two are on separate schedules.) Because Bulldozer-based CPUs are intended to be drop-in compatible with today's G34 and C32 Opteron sockets, AMD should have relatively little trouble persuading server makers to qualify and offer Bulldozer-based products when the time comes.
The outlines of the Bulldozer architecture haven't changed much since AMD revealed the first specifics last August. The chip is arranged around four "modules," each containing a pair of four-issue-wide integer execution cores and a single, shared floating-point unit. The FPU should be much more capable than a single FPU in today's Opterons, in part because it will include dual 128-bit execution units and in part because it will support both Intel's 256-bit AVX extensions and AMD's own 128-bit FMA instructions.
AMD has filled out a few details since then, including assigning the marketing name "Flex FP" to Bulldozer's floating-point unit. The implication of flexibility may be important in the upcoming skirmishes over FP math instruction sets and performance. We expect Bulldozer to reach about half of Sandy Bridge's peak throughput with AVX and 256-bit vectors, but with AMD's own FMA instructions, Bulldozer should offer peak rates comparable to Sandy Bridge's with AVX. Given the likelihood that Intel's ISA extensions will achieve broader adoption than AMD's, the prudent path for AMD may be to emphasize its flexibility and legacy compatibility over raw throughput. In fact, Detwiler told us he thought AMD would have an "advantage" because customers wouldn't have to use new instructions with Bulldozer.
Other bits of new info that have trickled forth from AMD regarding Bulldozer include an estimate of 50% higher compute and memory throughput than existing Opterons, despite an increase of only 33% in the raw core count. (The dual-chip Interlagos will have eight modules and 16 cores, while the single-chip Valencia will be a four-module/eight-core config.) Those gains will purportedly come from both higher memory frequencies and improvements in Bulldozer's memory controller. Also, Bulldozer's Turbo Core dynamic clock speed scaling will enable all 16 cores in an Interlagos processor to run at up to 500MHz beyond their base clock speeds. When fewer cores are active, even higher clock speeds are possible, though AMD hasn't divulged any further specifics about frequencies.