News - Intel shows off new eight core server CPU

[HEXUS]

The latest member of the Xeon family, codenamed ?Nehalem-EX? will start being produced later this year.

Read more.

[Michaelv]

sounds very exciting and useful for supercomputing needs such as simulations etc (virtualisation too)
but history has shown that when we get above 2 sockets, the price for the cpus themselves become simply ridiculous and in most cases it is far cheaper to purchase several dual socket servers. The trick here (and they know it) is that many applications cannot be clustered (in medical software, oil simulations etc) so big companies will pay BIG money for this...

[kalniel]

Thing is AMD had the >2 socket market pretty well to itself in previous years with much better scaling. They will be feeling sick looking at this chip.

[borandi]

sounds very exciting and useful for supercomputing needs such as simulations etc (virtualisation too)
but history has shown that when we get above 2 sockets, the price for the cpus themselves become simply ridiculous and in most cases it is far cheaper to purchase several dual socket servers. The trick here (and they know it) is that many applications cannot be clustered (in medical software, oil simulations etc) so big companies will pay BIG money for this...

Many of these applications are highly parallel, and benefit a whole lot more from GPGPU architecture upgrades than CPU upgrades. I did see a website heralding the 5870X2, to be released at some point this year, at 4.5TFLOP/s. It would take quite a few octacores to reach that level.

[antd]

8core! yummy indeed. i'll be next upgrading to an 8core

[Fraz]

Many of these applications are highly parallel, and benefit a whole lot more from GPGPU architecture upgrades than CPU upgrades. I did see a website heralding the 5870X2, to be released at some point this year, at 4.5TFLOP/s. It would take quite a few octacores to reach that level.

Sadly peak performance in GPGPU is nigh on impossible to attain. At an Nvidia seminar I recently went to, it took a lot of effort to reach 10-20% of the ~TFLOP/s peak performance in most cases. Only a very specialised set of problems is easily applicable to running on the nvidia GPUs. Worse still, you can divide those peak performance numbers by 8 if you want double precision calculations (at least until they put more double precision stream processes in the architecture).

In a lot of cases, you could argue that it costs a LOT more to pay someone to try and port the code over and get in running well than to just to buy some extra x86 servers... Obviously, porting the code is an investment, but depending on how much it'll be used, it may not be worth it.

One of the coolest things they did show us was a flow simulation (Navier Stokes) Airbus had done running via GPGPU. They had it working nearly 100 times faster than on a decent quad x86. Took a LOT of coding effort though.

[kalniel]

8core! yummy indeed. i'll be next upgrading to an 8core

What kind of server do you run?

[Moby-Dick]

What kind of server do you run?

A quad socket , 32 core box would make a very powerfull ESX host as long as you could afford sufficient memory for it - at a conservative overcomit of 200% thats 64 VM's