The Linux Beacon--Intel Announces First "Penryn" Xeon Processors

Newsletters	Subscriptions	Forums	Safari	Store	Career	Media Kit	About Us	Contact	Search	Home

Volume 4, Number 42 -- November 13, 2007

Intel Announces First "Penryn" Xeon Processors

Published: November 13, 2007

by Timothy Prickett Morgan

Chip maker Intel is timing the launch of its "Penryn" family of 45 nanometer quad-core Xeon processors to coincide with the Supercomputing Conference 2007 high performance computing event being held this week in Reno, Nevada, as a means of highlighting the performance and performance per watt advantages its new "Harperstown" Xeon 5400s can offer to HPC customers. But the chips will plug into other commercial two-socket servers and workstations, too, and they will put that much more pressure on Advanced Micro Devices and its "Barcelona" quad-core X64 chips.

In addition to the new quad-core chip, Intel is also expected to announce the "Wolfdale" dual-core Xeon 5200 series, which offers higher clock speeds on half as many instruction threads for customers who need maximum thread performance within the same thermals as the quad core chips deliver. Intel is also throwing a Core 2 Extreme variant of the Penryn family into the market for desktops and workstations.

Intel is the first chip maker to put the 45 nanometer chip-making processes into the market, just as it was first with 65 nanometer technologies nearly two years ago. The 45 nanometer processes use transistors that are based on a hafnium, high-k metal gate transistor technology that gives Intel a slight lead over IBM in the chip space, and probably a year over what AMD can bring to market.

The shrink from the 65 nanometer processes used to make the dual-core "Woodcrest" Xeon 5100s and the quad-core "Clovertown" Xeon 5300 processors to 45 nanometer processes is allowing Intel to cram a lot more transistors into the chips. Intel could jam 582 million transistors into the Clovertowns, but the 45 nanometer shrink with Harpertown allows it to boost that transistor count to 820 million transistors--all while staying in the same thermal envelope of the existing 65 nanometer chips. Most of those extra transistors are being used to crank up the L2 cache sizes on the chips by 50 percent, from 8 MB in the Xeon 5300s to 12 MB in the Xeon 5400s. (Having only one dual-core die per chip, the Woodcrest and Wolfdale chips obviously have half as much L2 cache memory per socket.) This extra cache memory is one of the ways that Intel boosts the performance of its processors as it has to keep clock speeds at or near 3 GHz for thermal reasons.

The other way that Intel is boosting performance with the Penryn chips is by ratcheting up the front side bus speed on the processors. The top-end Harpertown and Wolfdale parts have a 1.6 GHz FSB, and the regular parts have a 1.33 GHz FSB. The super-efficient, low-voltage Wolfdale chip--the E5205--runs at 1.86 GHz, has 6 MB of L2 cache, and has a 1.07 GHz FSB. By gearing this chip down, Intel can keep it running at a 65 watt thermal design point (TDP).

The regular Harpertown chips run at a low of 2 GHz and a high of 3 GHz; all of these chips have a TDP of 80 watts and only the two fastest Harpertowns--the E5472 at 3 GHz and the 5462 at 2.8 GHz--have the 1.6 GHz FSB; all other regular, 80-watt Harpertown parts have the slower 1.33 GHz FSB. Intel is also offering so-called extreme versions of the Harpertown chips, which have slightly faster clock speeds but also generate a lot more heat. The top-end X5482 has 1.6 GHz FSB and has a TDP of 120 watts. Dropping down to a 3 GHz part with a 1.6 GHz bus yields a 120 watt TDP. Two X5400 series Harpertown chips also have a 120 watt TDP, and run at 3 GHz and 2.16 GHz but with a 1.33 GHz FSB.

Aside from the low-volt dual-core Wolfdale E5205 chip, there are two other Wolfdales in the 16 chips that Intel has announced this week, and both are dubbed extreme editions. The X5272 runs at 3.4 GHz, has a 1.6 GHz FSB, and 6 MB of L2 cache; it has a TDP of 80 watts. The X5260 runs at 3.33 GHz and has a slower 1.33 GHz FSB, all in the same 80 watt envelope.

Intel's normal price list is missing from action as we go to press, but the company says that pricing on the Wolfdale and Harpertown chips range from $177 to $1,279 in the usual 1,000-unit quantities.

The Core 2 Extreme QX9650 code-named "Yorkfield," supports the 47 new SSE4 multimedia extensions that Intel has been talking about for some time, which helps with processing for games and other media-intensive applications. The chip runs at 3 GHz and has a 65 watt TDP. It costs $999 in 1,000-unit quantities. (These new instructions are also in the server variants of the chips.)

In terms of performance improvement, Intel is telling customers to expect servers running thread-friendly code such as Java applications that they can expect a Harpertown Xeon 5400 running at 3.16 GHz to have about 25 percent more oomph than a 3 GHz Clovertown. Obviously, the bus speed and larger L2 cache memory is a lot of that increase, but the extra clocks help, too. On the SPECint_rate2006 test, the Harpertown is showing a more modest 19 percent increase. On bandwidth-intensive applications, such as those used to model weather or financial transactions, performance clock-for-clock is from 40 percent to 50 percent higher comparing a top-end Woodcrest X5365 running at 3 GHz to a Harpertown 5400 running at the same 3 GHz. In this case, having twice as many cores, faster busses, and much more L2 cache is what helps performance.

Intel also said that it will ship low-voltage, quad-core Xeon 5400s with a 50-watt TDP and dual-core Xeon 5200s with a 40-watt TDP using the 45 nanometer processes some time in the first quarter of 2008. These low-wattage parts are definitely going to find their way into embedded devices, blade servers, and small form factor PCs.

The new Penryn processors are supported by the existing Intel 5000 series of chipsets for servers and workstations. Intel is also announcing new chipsets along with the Penryns, including the 5400 chipset (code-named "Stoakley") and optimized for HPC workloads, and the 3200 chipset (code-named "Garlow") that is designed for single-socket entry servers that use Xeon processors.