No mention of HP blades?
@article - The new architectures don't burn nearly as many watts as the old ones do, and DC power helps a bit as do the new Platinum power supplies. HP sells more blades than anybody, and has more bladesystem choices than anybody too. Modern processors and RAM are available in LV versions that don't take nearly as much power. Consider the maximum capacity of 4 10U C7000 chassis in a 42U rack, each populated with 16x BL2x220c G7 blades. Each of those blades is two servers. Each server can be configured with up to 48GB of LV RAM (Double when the LV 32GB DIMMS are validated) and, for example, a pair of the 60W 6-core Intel Xeon L5640 2.4GHz processor with hyperthreading and all the usual goodies. That's 4x16x2x2x6= 1536 Intel Xeon cores or 3082 threads in one rack. You're going to burn the industry standard 24K watts for the rack, yes, but you'll be able to get a lot for those watts too - VMHosting will oversubcribe those 4x at least, and serve hundreds or thousands of accounts on each shared VM.
Say in a web hosting VM such described you put 1000 accounts per core (500 per thread), with some software to migrate accounts to VMs to balance demand (most accounts do nothing at all, but the active ones need to be distributed between VMs to give good performance). One rack would be good for 1,536,000 accounts of which 7500 were very active. It would pretty much be paid for by subscriptions within the first week, yielding great profits from then on.
Modern datacenters are moving to 28K watts per rack. That C7000 solution is 15.26 Watts per core (and 4GB RAM per core) for the whole system. It's no competition for the K computer in Flops/W but it does a respectable job of hosting industry standard VMs without setting your rack on fire. It's almost as if HP and Intel are thinking a bit about this watts/rack issue - whodathunkit? There's even management controls so you can tell your C7000 not to burn more than 6KW or whatever limit you prefer, and it will downstep processors if it has to to prevent excursions.
At 160Gbps stock with FlexFabric or Flex10 interconnects, one single C7000 chassis exceeds the network bandwidth of Wikimedia, so upstream connectivity isn't an issue unless your switches ain't got the grunt. A rack of them is 640GBps of upstream bandwidth, which is - to be subtle, quite a bit. In that case each server has four "virtual NICs" to divvy up their bandwidth in 10Mb increments and if you're totally retro up to one virtual FC channel at up to 8GBps (depending on what's left after Ethernet is allocated. Most folks will allocate 4Gbits for FC and the rest to Ethernet) To enable FC on the blades of course FC uplinks from the interconnect is required, which diminishes net available Ethernet bandwidth - which is not in scarce supply so it's a small loss. If you enable FC, it's a one-hop deal - The FC port on your Flex Fabric interconnect has to connect to a FC switch. Routed or switched FCoE is an emerging standard that isn't fully baked, For now, DataCenter Ethernet is a one-hop deal.
@AC 22nd June 2011 17:22 GMT - Yes, All the new HP blades have internal USB and SDHC for booting from USB or SD. USB and SDHC are up to 32GB or more, so a respectable OS is possible but most people PXE boot or use ESXi, which doesn't need much. They can also come standard with 10G FlexFabric (FCoE), have Infiniband QDR and FusionIO based PCIe attached SSD storage available. They support PCIe attached sidecars if you're into the GPGPU thing, though that cuts your CPUs per rack in half. Most of the HP blades now support the new SFF SAS SSD's, that go up to 800GB each and really pump the IOPs but are quite spendy.
@everybody talking about thermals - If you feed the racks with outdoor ambient air and vent to stacks, then there is no thermal issue unless you chose for some reason to locate your datacenter in Phoenix, AZ. In some cases the delta between ambient and exhaust can be economically recaptured, or the hot exhaust can be used in other ways. There is no good reason for chilled datacenters. The servers work at 40C or 104F and the hotter they are, the better they work right up until they fail. Just put your servers where it doesn't get that hot and feed them filtered ambient air with a blower. For example if you're somewhere that it gets really freaking cold, put ducts under the sidewalks and parking lot and use your servers to heat the patio. If the load is light and the snow is deep, run benchmarks or fold some proteins. If you're somewhere that it gets hot, a heat pump with the cold side on the inlet between the filter and the server cold side should do on the rare occasions that it gets that hot if you're not like Google and can't shift load to a cooler datacenter - and it can be OFF the rest of the time. There is no good reason why I need to wear a sweater in your datacenter. That's just wasteful. Servers are not some odd alien species that thrives in cold air.