Overclocking post-Nehalem

Priorities change with Intel's new chip design.

Over the past few years, you would have noticed the emphasis on defying the gravity in the overclocking world. The "enthusiasts" were sometimes driving DDR2 memory up to 2.5 volts from its 1.8 volt defaults, or, more recently, DDR3 memory at beyond 2.1 volts from its 1.5 volt default - an utterly dangerous 40 per cent plus voltage jump in either case?

Not to mention the related North Bridge and then FSB overclocking and over voltage seen in such cases as well. Of course, the memory voltage would invariably end up much higher than the FSB or NB voltage either way.

However, with desktop Core i7 'Bloomfield' Nehalems around the corner - less than a week away - and AMD Deneb here as well a little later, the full-fledged North Bridge and FSB tuning will disappear from the high-end enthusiast checklist. Everything related to memory will be sharing the same die with the CPU itself now. To cut a long story short, are the priorities changing now?

As the new CPU voltages go below 1.2 volts, even if the voltages for the integrated memory controllers and QuickPath Interconnect or HyperTransport are decoupled, they still shouldn't differ drastically. Thus, on the first Nehalems we see an Intel recommended 1.65 volt limit for the DDR3, not much more than the basic 1.5 volt stock voltage for that DRAM kind.

All that said, in our chats with Intel we've discovered that the above 'maximum' really is just a recommendation, and that Intel itself has already pushed things as high as 1.8 volts.

Anyway, the first 'Nehalem-optimised' memory kits don't seem to suffer much from these restrictons - the Qimonda Xtune does DDR3-1900 CL9 at stock 1.5 volts, as we saw, and Kingston will offer DDR3-2000 CL9 tri-channel kits at the 'prescribed maximum' 1.65 volt settings.

Yes, this is not as good as DDR3-2000 CL8 setups you could achieve if going with some of the current DIMMs to near 2 volts, but it saves a bundle on power and especially cooling requirements - and it's more than compensated for by the lower latency integrated memory controller anyway.

A tri-channel DDR3-2000 setup will give you in excess of 48GB/s raw theoretical memory bandwidth, and easily around 30GB/s in tests like Sandra or Everest. That's good enough even once you (hopefully in exactly a year's time) upgrade your Core i7 Extreme setup to its 32nm 'Westmere' based 6-core 12MB L3 cache shrink, running somewhere above 3.6GHz in that same LGA1366 socket.

Talking about upgrades in that socket, do expect to see a few more Bloomfield steppings in the next few quarters. This is a brand new microarchitecture after all, and Intel will be tuning the stuff along the way.

A side benefit of the new platform is that, with the memory controller on the CPU, overclocking the North Bridge becomes far less important - in fact you can completely avoid it, unless you're trying to speed up the QPI from its (more than sufficient we'd say) 25.6GBytes/s bi-directional speed.