While Intel probably could have launched the server variants of its 64-bit “Nocona” Xeon processors a month ago when the chips and their related chipsets were made available for workstations, the company decided to hold off, most likely to give server makers time to qualify the chip for their designs and to give it something to say at the LinuxWorld show this week in San Francisco.
As has been the case with the past Xeon DP processors, the Nocona Xeon DPs used in servers are absolutely identical to the Noconas used in workstations. But the chipsets associated with the Noconas activate different features, depending on if they are used in a workstation or a server. The Noconas run at 2.8GHz, 3GHz, 3.2GHz, 3.4GHz, and 3.6 GHz, and they include 1 MB of on-chip L3 cache. They cost from $209 for the 2.8 GHz chip to $455 for the 3.2 GHz version to $851 for the 3.6 GHz version.
The Nocona chips have an 800 MHz frontside bus and obviously sport the Extended Memory 64-bit Technology (EM64T) that was embedded into the Prescott Pentium 4 core that Nocona and the future Cranford and Potomac Xeon MPs are all based upon. While the Tumwater workstation chipset used with Nocona workstations had beefed up graphics, the Lindehurst E7520 and E7320 chipsets for servers support a variant of the SpeedStep power saving and power management features that Intel created for its notebook and laptop processors called demand based switching. (In the Tumwater chips, this feature retained the SpeedStep name.) Like the Tumwaters, the Lindenhursts also support PCI-Express peripheral links and 400MHz DDR2 main memory. The Lindenhurst chipsets also have RAID 1 data mirroring inside the embedded disk controllers. The chipsets also support online spare memory, memory mirroring, and the failover of dual-channel controllers to single channel operation in the event one channel fails. Exactly what the difference is between the E7520 and E7320 chipsets is remains unclear, but a good guess is that the first supports five peripheral slots and the second three.
Adam Martin, enterprise marketing manager for Intel’s EMEA unit, says that the typical two-way Xeon server burns about 200 watts running the prior generation of Prestonia Xeon DPs at 3.2GHz , and that the average power consumption of a machine using 3.6GHz Nocona’s will drop to about 140 watts, a saving of 28% in electrical consumption and heat dissipation. He says that vendors will be rolling out Nocona servers through August, which is a slow time in Europe because of the lengthy vacations Europeans enjoy, for a big ramp in September. (Back in June, Intel said that an average workstation consumed about 400 watts and these similar Nocona technologies could drop consumption to about 300 watts. I think it unlikely that a workstation burns more juice than a server. Something is amiss in these comparisons.) As for performance, Martin says that on floating point workloads, the 3.6GHz Nocona servers will have about 30% more oomph than a 3.2GHz Prestonia box, while on Web workloads the performance gain will be even more–as much as 54%, as gauged by the WebBench benchmark test. Those comparisons are based on 32-bit mode operation.
Martin says that 64-bit commercial Linux distributions from Novell Inc and Red Hat can take advantage of the Xeon memory extensions, and while the delay in Microsoft Corp’s support for these extensions in an upcoming tweak of Windows Server 2003, which was expected this fall and is now expected sometime in the first half of 2005, is not a good thing, it should not hurt Nocona server sales. We have anticipated such delays, and it will have a minimal impact, says Martin. We are not expecting people to do an instant flip to 64-bit anyway. Having 64-bit hardware ready that can run in 32-bit mode as well–and do so with substantial performance improvements related to other features–is what will make Nocona servers a good buy compared to Prestonia units. If the 64-bit Windows operating system from Microsoft was ready and the hardware was not, this would be a worse scenario – provided customers care about 64-but support at all. Martin says that for most x86 applications, moving from 32-bit to 64-bit memory addressing will only boost performance somewhere between 8% and 10%. This is important, but not vital. The fact is, companies can buy the 64-bit hardware now, move to 64-bit Linux or Windows at their own pace, and then get down to the hard work of getting 64-bit applications running. The latter will take years.
As for the future of 64-bit Xeon MPs, Martin says that the Cranford Xeon DP, which is for streamlined four-way machines, is due sometime in early 2005. The Potomac Xeon MP will follow in the second quarter of 2005 and is aimed at bigger four-way, eight-way, and even larger SMP servers. Potomac is expected to have a larger on-chip cache than Cranford; how else it might differ, Intel is not saying.