AMD vs Intel: Computex Round One, Servers Get Attention

Although Intel’s now-defunct NetBurst microarchitecture’s poor IPC (instructions-per-clock) characteristics on conventional code sequences were undesirable to a greater or lesser degree across all computing segments, they were especially problematic in particularly power consumption-sensitive applications. NetBurst’s prodigious electron appetite is why Intel focused its initial resurrection attention on mobile computing with the Pentium M CPUs and their revisited P6 microarchitecture. And it’s also why, particularly until follow-on Core microarchitecture-based Woodcrest server-tailored chips hit the market, AMD happily found itself in a notably strong competitive position with its Athlon-derived Opteron CPUs.

Woodcrest is nearly three years old at this point and has been superseded by several descendents; as such Intel’s regained some of its historical swagger in the enterprise. However, ‘big iron’ designs (therefore design-ins and -outs) take much longer than in consumer-targeted product categories, thereby giving AMD a notable stable business bright spot as the decade has progressed. IT sales volumes are comparatively low versus the consumer market, but they’re balanced against atypically healthy server profit margins. As such, recent announcements from both companies have captured my attention.

Back in September of last year, Intel formally unveiled (with plenty of prior pre-announcements) its Dunnington CPU (the Xeon 7400 series), derived from the 45 nm variant of the Core microarchitecture (codenamed Penryn). Dunnington is a six-core-per-die design, with no HyperThreading support. Its production release roughly coincided with the launch of Core i7, the high-end consumer initial spin of Intel’s next microarchitecture, Nehalem, also initially fabricated on a 45 nm process lithography. With past history as a guide to future trends, it wasn’t much of a stretch to predict that a Nehalem-derived server CPU was on the way…

…and in fact Intel unveiled it in two-punch form. Nehalem EP (i.e. the Xeon 5500 series), intended for one- and two-CPU systems, began shipping several months back. Each Nehalem EP CPU contains two or four cores per sliver of silicon along with per-core HyperThreading support, translating to four or eight virtual cores per die. And for higher-end system designs containing four or more CPU sockets, Intel last week pre-announced its Nehalem EX CPU, with eight physical cores per die (once again with per-core HyperThreading support). Thereby begging the question of what if any meaningful role the Itanium CPU has in Intel’s arsenal going forward…

Note that Nehalem EX won’t be formally unveiled until some time in the second half of this year, and that Intel doesn’t anticipate its customers will start shipping systems based on the chip until some time in the first half of next year. So why did the company conduct such a premature public leak last week? One word, I ’spect: AMD. Today, Intel’s pugnacious primary competitor rolled out a six-physical-core, 45 nm-fabricated chip code-named Istanbul and formally named the Opteron 2435 in its initial incarnation. Istanbul is an optical extrapolation of the existing four-core Shanghai CPU and a descendent of the troubled ‘Barcelona’ quad-core Opteron, with all products based on the K10 microarchitecture.

In welcome news for AMD, Istanbul is arriving nearly a half-year ahead of its initially published schedule. It’s good to see that the company seems to have regained its design stride after a disastrous Barcelona-and-Phenom 65 nm stumble (though 45 nm yield robustness is still yet to be proven). AMD claims that Istanbul is already shipping for revenue to key customers. And given that Istanbul supports up to eight-way CPU system designs, the processor’s availability gives AMD a roughly six month lead in such complex configurations over Intel’s Nehalem EX.