When it comes to servers, consolidation and virtualization are the two trends front and center on everyone's radar screens. Yet ongoing advances in processor performance often get short shrift -- perhaps because they're taken for granted. Now, AMD and Intel are kicking their respective architectures up another notch. The upshot, I believe, is that we are about to open a new chapter in the reinvention of the data center.
|(Click for larger image and to see more slides.)|
Why? Well, if virtualization is the icing on the cake -- because it enables blades to be sliced and diced into a dizzying number of logical processors -- then the processors themselves are the cake. You can't layer on additional application and operating-system instances and/or threads without the raw CPU horsepower to support such throughput. Nor does it make sense to address infrastructure virtualization until you've successfully tackled it on the server level.
Yet what the focus on virtualization makes clear is that hardware is now so powerful -- I've argued that computer cycles are now essentially free -- that the processor is almost the least of the worries of data-center architects (except insofar as one has to pay for electricity and cooling).
True, there's still lots more efficiency that can be squeezed out of software virtualization, as currently deployed. Which is precisely where I'm headed with my argument. Namely, that the new crop of more powerful processors will allow CIOs and data-center designers to place their focus squarely on these critical elements:
Before I go overboard and argue that the coming crop of processors can also write their own code, let's look at the products themselves.
AMD's server-processor attack is dual-pronged. On the one hand, the scrappy semiconductor vendor is boosting the number of physical cores available in a single package. At the same time, it's working hard to ratchet down power dissipation, consistent with its ongoing marketing message of "efficiency computing."
The latest generation to enter the AMD Opteron server family is the 6-core Istanbul, introduced earlier this summer. According to AMD, the new parts deliver 30% more performance than the existing quad-core Opterons, while staying within the same power envelope. (That would be 55W, 75W, or 105W, depending on which speed grade you're talking about.
In an additional green push, AMD followed up its initial Istanbul intro with a 40W part. AMD says this device boosts performance-per-watt by 31% compared to quad Opterons. The downside is that the power savings are achieved by scaling the clock speed back to 1.9-GHz. (That compares to the 105W dissipated by the hottest and fastest Istanbul, clocked at 2.8 GHz.)
Moving up a notch in the more-cores race is Magny-Cours. That's the code name -- after a French Formula One Grand Prix site -- for AMD's 12-core server processor, which was shown in August at the annual Hot Chips conference. The chips will ship in 2010.
But Magny-Core is about more than simply upping the core complement. It will debut the 2.0 iteration of AMD's Direct Connect Architecture. Along with the raw metrics of a 3.3x boost in memory speed and 1.9x the HyperTransport bandwidth, AMD says this means it'll run more virtual machines per server.
Regarding the dual trends of boosting the number of cores while trying to maintain the thermal envelope (i.e., not boosting power dissipation), AMD has two lines in one of its PowerPoints which I thought rendering the issue very succinctly:
"Virtualization drives the need for more cores and greater stability.
Cloud computing and dense deployments drive the need for greater energy efficiency."