Quote:
|
Originally Posted by Butcher
SMP boxes do not have their own memory per chip... They use the same memory and the same memory bus. This is why cache coherency and similar issues become such a problem. I don't see why a dual core is any worse than SMP, if anything it's better because it has the potential for a high bandwidth path between the CPUs as in AMD's solution.
|
Actually they do
Its called NUMA, it stands for Non-Uniform Memory Access.
AMD's Opteron platform uses this. If you look at a dual (or quad) Opteron board you'll notice that each CPU has its own memory bank, so CPU1 doesn't have to go through CPU0 to access the memory. The memory banks and respective CPU's are connected together via hypertransport, which means a dual Opteron system has more than 11,000mb/sec memory bandwidth available.
On the new nForce4 Opteron boards (well the ones that use both the Pro2200 and 2050 chipsets), you effectively get two machines on one board. Each chip has its own chipset, memory bank, PCI-E 16x slot (with full 16 PCI-E lanes, not 8/8), and its own gigabit NIC. On top of that there's the AMD8131 PCI-X chip.
And on top of all of that its dual core ready
Quite a few of reviews for dualcore chips have been tested on nForce4 Opteron platforms, ie the Tyan Thunder K8WE.
Whats even better is that these boards have nvidia nTune support for software OC'ing, but after lots of research a bunch of us have found out ways to OC the hell of out of the platform via BIOS and a few board mods
1.8ghz 244's running at almost 2.2ghz on stock voltage. As soon as we sort the voltage out, (and with the proper cooling of course) we'll be able to push these quite far.
Whether the dual-cores will clock as well is another story.
I plan on buying a pair or 244 E stepping Opterons until next year when I'll upgrade to dual core.