NEC Launches Two-Way Itanium Blade Server

by Timothy Prickett Morgan

NEC is a long-time partner and rival of Hewlett-Packard, and the two companies have both placed substantial bets on the Intel Itanium architecture. Despite all of the noise about Advanced Micro Devices' Opteron processor and the advent of 64-bit Xeons from Intel, NEC, like HP, remains committed to the Itanium processor, and not just at the high end of its server line. This week, NEC is rolling out its third generation of blade servers, and this one, called the IPF Blade, is based on Itanium.

The announcement makes NEC the first company to ship Itanium blade servers in volume. While IBM and HP have dominated blade shipments to date, most of their blades--in terms of shipments--are based on Intel's Xeon processors. IBM will this week get its PowerPC 970-based blades out the door, and HP is flirting with Opteron and Itanium blades. But for the lowest cost per flops on technical workloads, a compelling case can be made for Itanium, and that is what NEC intends to do.

In the IT market, NEC is perhaps best known for its first generation AzuzA and second generation Asama Express5800 32-way Itanium 2 servers, which are among the most scalable servers in the world. These are general purpose machines that can run Windows and Linux--and even HP-UX if NEC wants to, since NEC has been one of HP's Unix development partners for a long time.

In the blade market, according to Scott Schweitzer, product manager for the Itanium server family with NEC Solutions America, NEC will be focusing on peddling Itanium blade servers into the clustered Linux environment, particularly for high performance computing (HPC) customers who want high-speed InfiniBand interconnections as the backbone in their clusters.

While NEC is announcing the IPF Blade this week, the server will not start shipping until September, since NEC timed its development and manufacturing for the IPF Blade to the delivery of the Madison Itanium 2 processor with 9 MB of cache. While this chip is expected to be available at 1.6 GHz, 1.7 GHz, and 1.8 GHz clock speeds, according to the industry scuttlebutt, NEC is only planning on using the 1.6 GHz version of the chip in its blade servers because the extra flops in performance is not worth the extra heat they generate.

A fully loaded IFP Blade chassis, which has nine blades with two 1.6 GHz Madison, 9 MB chips per blade, will consume approximately 5,700 watts of electricity as it runs, delivering about 100 gigaflops of computing power. Three of these IPF Blade chassis can be packed into a standard 42U rack, with enough space left over for interconnection and storage, for a total of 300 gigaflops per rack. Each blade has 12 DIMM memory slots--for a maximum of 24 GB of main memory, soon to be doubled to 48 GB--a single PCI-X slot, and dual Gigabit Ethernet ports.

For those who want a low-latency, high-bandwidth interconnect, NEC has partnered with TopSpin Communications for InfiniBand interconnection switches. NEC is using TopSpin's 10 Gigabit (4X) InfiniBand interconnect today, and will move to the 30 Gigabit (12X) InfiniBand. The addition of faster and wider interconnection will allow thousands of blades to be linked together into a single cluster. The blades in the IPF Blade chassis are hot-pluggable, and they include space for two 36 GB or 73 GB SCSI disk drives. However, with the boot-over-InfiniBand capabilities of the TopSpin switches, many HPC customers will probably boot over the SAN rather than store OS instances locally.

Schweitzer says that NEC is going to try to sell the chassis with nine blades--presumably with a single Itanium 2 processor--2GB of main memory, a 36 GB disk on each blade, plus redundant power supplies, installation, systems management software, and maintenance, for about $79,000 per chassis. If you do the math, that works out to around $8,700 per blade. NEC is going to sell these directly around the world as well as through its distribution channel.

On the software front, NEC will also be delivering its own Deployment Center operating system provisioning software with the box, and has similarly created its own blade management program. Ironically, NEC is only supporting Red Hat Advanced Workstation on the server, since the feature set and price tag on Enterprise Linux 3.0 is way too high given the number of nodes HPC customers normally use. The HPC shops basically want a kernel, some libraries, compilers, and tech support. Hence NEC smartly is not supporting the full-blown Linux distributions from Red Hat. Schweitzer says that NEC will eventually support a variant of Novell 's SuSE Linux, but would not say which one. A version of the new SuSE 9 Professional tweaked for the HPC customers seems likely.

As for Windows support on HPC clusters, that will not be happening this year, but if Microsoft pushes into the HPC market in a big way, it could happen next year. NEC is looking at third parties--the most obvious are Platform Computing and the Scyld subsidiary of Penguin Computing--to supply Linux clustering support to link all of this together. Schweitzer says that NEC needs to keep its options open, since the HPC market uses a wide variety of clustering technologies.

While NEC is primarily targeting the IPF Blade at the Linux-based HPC clustering customers--meaning those who do simulations, financial modeling, genomics, weather modeling, weapons research, and similar applications--there is obviously a nascent market for clustered databases using Oracle's 9i RAC and Oracle 10g and IBM's DB2 ICE. NEC has been demonstrating IBM's future Stinger DB2 database on the Asama NEC 1000 enterprise servers, and it can quickly add parallel commercial database as a second market to chase as it develops further.