Intel Versus AMD: Profits Versus Cores-For-Free

Earlier this week, yet another skirmish in the seemingly never-ending AMD-vs-Intel processor war erupted. On Monday, AMD unveiled its eight- and twelve-core (dual-die in both cases) Opteron 6000 Series CPUs. Some writeups also discussed a four- and six-core (single-die) upcoming variant known as the Opteron 4000, perhaps obviously with half the memory channels of its Opterion 6000 bigger sibling; out of respect for news embargoes, I won’t comment. The very next day, Intel wrapped up yet another multi-stage product rollout, formally launching the Nehalem microarchitecture-based four-, six- or eight-core (single-die in all cases) Xeon 6500 and 7500 Series CPUs (which, if your application can take advantage of HyperThreading, deliver an additional four, siz or eight virtual cores).

The Xeon 7500 (aka Nehalem-EX), manufactured on Intel’s 45 nm lithography process (not the company’s latest 32 nm process) comprises 2.3 billion transistors and runs at up to 2.66 GHz (6-core) or 2.26 GHz (eight-core) clock speeds, with integrated L3 cache sizes of up to 24 MB. AMD’s dual-die ‘Magny-Cours’ beast, conversely, has a cumulative 692 mm² die size (nearly 2 billion transistors), encompassing 19.6MB of L3 cache, and runs at clock speeds of up to 2.3 GHz. It does so at a notable per-core power consumption advantage over the Intel counterpart, but keep in mind that this is only comparing CPUs; chipsets, DRAM arrays and other necessary processor companions can blur overall distinctions.

‘Magny-Cours’ is an enhanced version of the prior 45 nm Istanbul processor for servers, tailored for 4P and higher-end configurations. It supports DDR3-1333 SDRAM, boosting overall performance in certain memory bandwidth-bound applications. It also devotes a portion of each core’s L3 cache for use as a ’snoop’ lookup table (an approach known as HT Assist), intended to minimize bandwidth-consuming core-to-core communication over the HyperTransport v3 links that interconnect them. Intel’s latest CPUs, on the other hand are 4P-tailored versions of the earlier-unveiled Nehalem-EP (45 nm) and Westmere-EP (32 nm) processors, and their presence on the company’s product line puts the future of the Itanium CPU in further doubt.

How do the products stack up? From a performance standpoint, I haven’t yet seen any Nehalem-EX versus Opteron 6000 benchmark shootouts; the closest study I’ve come across is one that AnandTech published a few days back, which stacks up Opteron 6000 versus Westmere-EP. Keep in mind that in this particular case, what are being compared are a two-die, eight- or twelve-core AMD CPU manufactured on a 45 nm process versus a six-physical-core (twelve-virtual-core) Intel processor built on a 32 nm lithography, therefore with increased clock speed up to 3.3 GHz versus the Nehalem-EX. AnandTech’s results are perhaps not surprising; substantially multi-threaded software takes significant advantage of the AMD physical core count advantage, whereas less core usage-efficient applications (or, conversely, applications that can more efficiently tap into HyperThreading-enabled additional virtual cores) leverage the Intel CPU per-core clock speed lead. Nonetheless, I offer it to you for perusal and extrapolation to Nehalem-EX and your particular design situation.

What of pricing? Here’s where, as usual, I pause and take a deep breath when I look at the data. Here, courtesy of The Register, are Intel’s Nehalem-EX published budgetary prices:

And on the far right of the below graphic (again sourced from El Reg) are AMD’s pricing plans for Opteron 6000:

Divide the price tags by the core counts and you’ll likely discern, as I did, that AMD’s got quite a ‘fire sale’ underway. Either Intel’s making a ridiculous profit on its products, AMD’s making little to none, or some combination of both factors is at play. As usual, AMD is substantially undercutting Intel-equivalent CPUs in order to secure sufficient market share.

I will give AMD credit for at least one intriguing fiscal twist; since server operating systems are typically sold at prices that depend on the number of per-system CPUs they’ll be run on (measured by number of packaged units, not number of cores per package), the Opteron 6000 core-per-packaged-CPU advantage likely results in lower operating costs for elaborate system configurations.