A cluster is a bunch of separate computers that communicate by IO and messages. A NUMA system's big big thing is to maintain cache coherence. Which is it? Sequent's definitely NUMA, not cluster. But optical links? I associate those with latency that's way bigger than desirable for memory accesses (orders of magnitude lower than network latencies).

What is meant in this article by the word 'cell'?

I suspect it does not mean IBM's Cell CPU...

According to the spec sheet for the Unisys ES7600R, a cell is the basic rack-mountable unit you buy -- each has up to 16 processors (presumably with 4-core chips), its own memory, and IO. Bigger systems are made by gluing multiple of these together. Not small. Each is 7" high, rack-wide and deep, weighs 95 lbs., and consumes 0.77 KVA.

Apparently, by the way, to answer my own question above, the glue connecting them is memory access spanning cells: A processor in one cell can access memory in other cells. And do IO to devices attached to other cells. So it's NUMA, not a cluster. But with that optical interconnect, there must be a big latency hit going to other cells.

HP Superdome scales to 16 cells and 64 sockets (that's 128 cores with the current dual-core CPUs, and 256 cores when Tukzilla hits next year), without the need for fiddly optical interconnects, just high-speed and wide-bandwidth crossbar switch backplane. So Itanium scales further and offers far better performance, and uses PCI-e or PCI-X card cages for greater interconnect options, including Infiniband and 10GbE.

HP VSE under hp-ux can already host virtualised Windows Server images as separate OS instances (along with Linux, VMS, and other hp-ux instances). Or, you can use harf partition for completely electrically isolated OS instances in the same frame and cluster either inside the frame or between frames, which is better resilience than software virtualisation.

The HP Integrity server range scales from a 2-socket blade via 2-, 4-, 8-, 16-, 32- to 64-sockets, which is much more choice.

HP Integrity comes with five-nines availability. If you want to go further you can go Integrity NonStop (also Itanium), with which some people I know claim they never have any downtime.

Even the FSC Itanium range has more to offer! This is just a cheap alternative to low-end Itanium, and existing large Xeon servers such as the 8-socket ProLiant DL765 would seem to offer a much cheaper and better tested means of hosting a large Windows virtualised instance, with two such servers offering better redundancy too. What happens when Dell, HP or IBM start putting six-core Xeons into their large x86 servers - bang goes more than half of the NEC/Unisys market. I know NEC and Unisys used to have a rep in the datacenter, but how long ago was that?

adding numa on top of the fsb does not sound like the cleanest approach... monster is quite an apropriate name...

mr. Matt , the Itanium might scale well (for the few apps that it runs), but these Unisys systems are for people who want to take advantage of binary compatibility, and scale up their X86 apps without putting effort to migrate them... Linux and Windows ru on x86 for quite a while...

Both Sparc and Power architectures have had binary compatibility for decades and scale better today than Itanium systems (256 Sparc cores in the M9000 today).

Itanium will remain a niche product, will have to compete not only against X86, but also against sparc and power systems ...