Intel’s 45nm Penryn\/Yorkfield architecture packs serious punch

Early this morning, NDA restrictions lifted on Intel's next-generation 45nm Core 2 Extreme QX9650 Yorkfield/Penryn CPU. Reviews hit the Internet minutes later, treating those of us who've been curious about Penryn's performance to plenty of information on the new processor. There are any number of QX9650 reviews out there—I'd recommend Michael Schuette's over at Lost Circuits or Scott Wasson's (tech) report. Despite the NDA lift, the QX9650 isn't expected to be available for purchase until November 12, and Intel has yet to disclose any pricing information. Given the "Extreme" moniker, however, I'd expect the chip to debut around the $999 price point. HangZhou Night Net

As for raw performance, Yorkfield typically streaks past the QX6850 (its 3GHz Conroe-based quad-core predecessor), while drawing one-third the power of its counterpart at idle and just under half at load. The processor's performance improvements and dramatic power reductions are delivered by a number of architectural advances Intel implemented in its 45nm process technology, including hybrid gates that incorporate hafnium, 12MB of 24-way set associative L2 cache (up from 8MB), SSE4 instructions, and and several other design improvements. If you'd like more information on Penryn and its upcoming successor Nehalem, check Jon's coverage.

In terms of overclocking potential, Yorkfield seems to have quite a bit of headroom. Opinions on exactly how much vary from review to review, but it seems that a 3GHz QX9650 is easily capable of hitting 3.6GHz, often managing 4GHz (sometimes with a voltage increase, sometimes not), and has an upper range (on air) of between 4.4 and 4.6GHz. Obviously there are no guarantees when it comes to overclocking, but Intel likely has the ability to boost its 45nm speed grades virtually at will.

As with all new instruction sets, existing code must be rewritten to incorporate Intel's new SSE4 standard. Currently, support appears limited to VirtualDub 1.71 combined with DivX 6.7. Even then, the program's code needs some fine tuning. As Michael Schuette covers at Lost Circuits, enabling SSE4 natively with multithreading also turned on does boost performance significantly compared to performing the same operations using the SSE2 instructions, but it doesn't reach the baseline established by disabling SSE4 entirely. The difference is small, but apparently consistent; the QX9650 without SSE4 activated is slightly faster than the QX9650 running with SSE4 on. This is probably more an issue of proper software support then it is an indication of weak functionality in the SSE4 instruction set.

As for what Penryn's dramatic debut means for AMD's rumored Phenom launch later in November, the short-term answer is "not much," given that AMD doesn't currently compete in the $1K market segment Yorkfield is likely to target. That gives Sunnyvale a little more time to ramp Phenom production, but the clock is running out. Intel is expected to introduce dual and quad-core desktop Penryn parts early in the first quarter of 2008, with Harpertown-based Xeon parts following as well.

As we've discussed before, there's reason to believe AMD may be having trouble ramping Barcelona/Phenom production. If that's actually the case, the company is in a sticky situation. Delaying Phenom's launch even further would give Sunnyvale more time to improve yields and launch clockspeeds, but is highly unlikely to play well with investors. On the other hand, AMD could paper-launch Phenom in "limited availability" mode, but company's ability to do this without damaging the new Phenom brand name is limited. If Phenom performance is excellent at 2.6GHz+, consumers might be willing to wait for the product, but they won't wait forever. Once desktop Yorkfield parts arrive, AMD's claim to performance-per-watt superiority will vanish, unless Phenom demonstrates significantly lower power consumption on its own. In addition, AMD will have to position its new architecture aggressively enough (if possible) to demonstrate that the company has achieved rough performance parity in at least some areas, without pushing Intel so hard that the company elects to jump into another price war.

Ultimately, Phenom needs to be ready to slug it out with Yorkfield desktop parts by the time the latter launches in 2008. If it can't, AMD is likely to be pushed back into the same value segment the company occupied in the K6/K6-2 days. AMD has bounced back from a number of similar scenarios before, but hasn't compared this poorly to overall Intel performance since the K6-2 versus Pentium III days of yore. Hopefully Phenom's launch won't mark the end of nine years of progress.

If you'd like a preview on the various upcoming and already-launched architectures, check the table below.

CodenameYorkfieldKentsfield (Conroe)WolfdalePhenomProduct seriesCore 2 Extreme 9650Core 2 Extreme 6850UnknownPhenom X2, X4Market segmentHigh-end enthusiastHigh-end enthusiastDesktopDesktopLaunch date11/12/2007Now1Q 20083Q 2007Cache size12MB (L2)8MB (L2)6MB (L2)512K L2 per core, 2MB L3 sharedProcess technology45nm65nm45nm65nmNo. of cores4424FSB1333MHz1333MHz1333MHzn/aClockspeed3GHz3GHzUnknown, assumed to be 2-3GHzUnknown, assumed to be 2.0-2.4GHz