So Intel finally got it together and released Nehalem today. It's one hell of a CPU, that's for sure. Frankly, if the performance increases present in the Nehalem CPU had been widely known prior to today, nobody would have bought Harpertown-based servers. They're effectively orphans as of right now for general server computing needs. The world moves on.
But as I said in the feature I wrote for today's launch, where does CPU development go from here? The punch Nehalem packs is essentially AMD's Opteron blueprint writ large with the substantial addition of DDR3 RAM. They took the NUMA architecture found in the Opteron and made it faster. But you can only use that particular potion once.
[ InfoWorld's Test Center has more on Intel's powerful new processor: Intel's Nehalem simply sizzles ]
There's no mistaking the fact that AMD has to step up to the plate here. It needs to deliver a Nehalem-killer -- and pretty quickly. Not only does it need to meet or beat the performance numbers put up by Intel, but the company must equal them in core development and really hammer the six- and eight-core chips that can compete core-for-core with the Nehalem's performance. I honestly have no concrete idea if AMD can pull that off. To be honest, I was fairly surprised that Intel was able to produce such a powerhouse at this point in time, and AMD doesn't have the deep pockets that Intel does.
Predictions for processors
So what's next for processors beyond Nehalem? More cores per die, more integration with previously outboard components, and more transistors. 32nm will rule the day, but when is the next time we'll see a huge performance boost in a single release like this? If Nick Knupffer is right and this is the biggest step Intel's made since the Pentium Pro, then will it be another 14 years before we make another leap of this kind? I doubt it. I think it'll all happen much quicker than that.
Perhaps we'll finally see the demise of the general-purpose CPU, and advances in CPU development will follow specific use cases and result in purpose-built servers that are task-specific. This would seem to be the model that Sun's following with the SPARC, foregoing single-thread performance for multiple threads per core and multiple cores per CPU. In a smaller way, AMD and Intel are beating that drum too, but not quite as hard.
It really all comes down to software -- if developers aren't busy trying to properly thread their code, they're going to get left in the dust.
But all that remains to be seen. I've read estimates that claim somewhere on the order of 40 percent of installed servers are still single-core. Especially with the world-wide economy on shaky ground, it'll take some time before those boxes are mothballed. And they'll be replaced by virtualization frameworks, no doubt about that -- virtualization frameworks running on four-, six-, and eight-core CPUs.
Today is a good day for Intel, a good day for IT, and a good day for technology in general. As long as we can hold off the approaching Thunderdome, the next few years should be even better.