IBM Delivers Nocona Blades, Readies Opteron Blades
by Timothy Prickett Morgan
IBM is trying to keep the heat on its rivals in the blade server business by getting the coolest chips possible into its blades. To that end, the company has announced that the low-voltage variants of the "Nocona" Xeon DP processors have been certified in the blades that go in its BladeCenter chassis. The company also announced a new switch from Nortel Networks, and the top xSeries brass are hinting at plans for Opteron-based blades this year.
By moving to the low-voltage variants of the Nocona chips in the HS20 blade servers, IBM says that it can offer the full performance of the Xeon processor, which you will remember is Intel's first and only 64-bit X86-compatible processor so far, and cut a little more than 50 watts of power consumption per processor in the BladeCenter chassis. On a per-chassis basis, with a fully populated BladeCenter with 14 two-way blades, that works out to about 1,500 watts. This may not seem like a lot until you have to pay the electric bill for the chips and then the air conditioning to cool them; when you add those savings up across dozens or hundreds of servers, it becomes substantial. IBM is planning on shipping a 2.8 GHz version of the low-voltage Nocona starting at the end February; IBM is currently shipping 32-bit Xeon DPs running at 2.8 GHz and 3.06 GHz in the HS20s, as well as 3.2 GHz and 3.6 GHz Noconas running at the regular, higher voltages.
One lesson from the early days of the blade server market (and one that was, quite frankly, somewhat surprising) is that vendors did not want to sacrifice performance in order to save on electricity or cooling costs, or to save on floor space and get the most density. Hewlett-Packard's early "QuickBlade" BL10e blades could pack 20 of Intel's low-voltage Pentium III or Pentium 4 processors in a chassis, and for running Web infrastructure workloads these machines had plenty of computing power. But they did not have anything near the performance of a Xeon processor, they did not have links into SAN architectures, and they did not have significant memory expansion. So companies did not storm to the low-voltage, low-power blades that HP was initially selling, and IBM never created a blade akin to the BL10e. The major selling points of blades--density and power efficiency--were not as important to early adopters as a bit more density compared with 1U rack-mounted servers, less cabling, and the prospect of better management tools (the latter being a dream that is still, so some IT shops with blade servers tell us, far from being realized.)
For those of you who are network junkies, IBM has also announced a second integrated switch from partner Nortel, which launched the first integrated switch for the BladeCenters two years ago. (Both IBM< and HP now have switches from Cisco Systems available as alternatives to the Nortel switches.) The new Nortel Layer 2/3 Gigabit Ethernet Switch Module, as the new switch is called, can consolidate and manage both Layer 2 and Layer 3 of the TCP/IP protocol. (The former controls switching, while the latter controls routing.) IBM says that this is the first integrated switch/router available in a blade form factor, and that the device will be sold exclusively though IBM and its channel partners, with service and support being handled by Nortel technicians.
On the BladeCenter alliance front, whereby IBM and Intel have released the specifications for all but the chassis and processor blades used in the BladeCenter and equivalent white-box Intel blade servers, IBM says it now has 150 vendors that have licensed the specs on a royalty-free basis. Host-bus-adapter maker Emulex has announced a special Fibre Channel adapter specifically for the BladeCenter called the LP1005DC, thanks to the licensing of the specs.
While all of this gadgetry is interesting, what blade server customers (and particularly those in the high-performance-computing market) are seemingly most interested in getting is the most computing power per watt. And to that end, Susan Whitney, general manager of the xSeries X86 server unit at IBM, says that the company will deliver blade servers based on the future dual-core Opteron processors. Advanced Micro Devices is expected to deliver its first dual-core Opterons beginning in the middle of this year, and it is likely that IBM's blades will use both the "Denmark" (100 Series) and the "Italy" (200 Series) for single-socket and dual-socket blades with two or four processor cores. It's remotely possible that IBM could go whole hog and make a four-socket blade, as it does for the current Xeon MP HS40 blades, and create a stunning eight-core, four-socket blade based on the "Egypt" (800 Series) future Opterons. Whitney was not specific about IBM's plans, other than to say that IBM would be "time to market" on dual-core Opteron blades.
She also indicated that IBM was looking at "hosted clients," which is IBMspeak for using a skinny blade server as a centrally managed PC that is remotely accessed by users, mimicking what HP eventually did with the Quickblade BL10e blades, and somewhat like what PC blade vendor ClearCube has created. She said that IBM and its partners are working on iSCSI connectivity, an InfiniBand switch, new host bus adapters, and a 4 Gbps Fibre Channel switch, too.