Intel to chase clouds with custom mobo
More Nehalem goodies
For a little more than a year now, IBM and Dell have been offering custom server designs to hyperscale data center operators that deliver more oomph for a given amount of juice. And it looks like Intel is going to be creating some specialized motherboards for the future "Nehalem" family of Xeon processors so it can get in on the hyperscale action.
As we reported earlier this week, Intel has hitched itself to the cloudwagon and reckons that hyperscale data centers - like those ran by Google, Microsoft, Yahoo, Amazon, Facebook, Salesforce, and so on - will account for anywhere from 20 to 25 per cent of its server chip sales by 2012. At Dell, custom servers sold to such customers already account for a significant share of shipments - at least 10 per cent and maybe as high as 20 per cent, by our estimates.
Intel estimates that as much as 14 per cent of its current server chip shipments go into such big data centers. (That may say more about how SMB server sales are drying up than it does about hyperscale sales these days). As Dell's experience - and the lack of sales into the Google account - shows, such customers want customized products that push the limits of efficiency and performance, and they don't want to pay much of a premium for the innovation either.
Intel wants to help facilitate the hyperscale build-out and the cloud computing revolution and to that end the company is preparing a new energy-efficient motherboard, called "Willowbrook," for its impending "Nehalem" family of processors.
Jason Waxman, general manager of high density computing within Intel's Server Platforms Group and the executive leading the cloud charge, says that a typical two-socket server used in a cloud has a 250-watt power supply and that it draws maybe 200 watts of juice. With the future custom Nehalem motherboard, Intel will use a slightly different layout for components that "unshadows" components so they are not stacked one behind the other, allowing the board to be cooled more efficiently.
This will also increase the efficiency of voltage regulation so the idle power consumption of a system created using the custom mobo will be under 85 watts. Waxman said that the standard Nehalem boards - and presumably be was talking about two-socket boards based on the "Tylersburg" chipset using the four-core "Nehalem-EP" processors set to launch any day now - burn somewhere between 110 and 115 watts when they are idling.
Waxman did not provide details of the Willowbrook motherboard design, and Intel probably won't say much about it until the Nehalem processor launch. He also did not say if the Willowbrook board was for rack or blade severs. But if Intel is making a custom board for rack servers, it will probably want to do one for blade servers as well.
The unshadowing of components seems simple enough in concept, but proved difficult with the chipset and front side bus architecture of current Xeons. The integrated memory controller on Nehalem chips and the simplified chipset means that components on the board can be moved around so they don't obstruct airflow. At the SC08 supercomputer show back in November, Silicon Graphics was showing off a Nehalem blade board (made by Super Micro, as it turns out).
Sun Microsystems is putting two two-socket Nehalem boards on a single blade, and this Sun blade could be based on the Willowbrook mobo. (The Sun Blade board was not from Super Micro, and it is very likely an Intel board, given Sun's partnership with Intel and its desire to not spend engineering money on anything it doesn't have to).
It will be interesting to see if Intel can reduce the idle power of the motherboard without having to resort to using low-voltage processors. Low-voltage processors have been available for years, which consume less power at peak performance but which sacrifice a fair amount of clock speed and performance to keep their thermal profiles low. While such low-voltage parts have been adopted by certain customers, most were not willing to pay a premium for a slower, albeit more energy efficient, part - even if that chip would pay for itself in electricity savings over the course of a year or two.
In hyperscale environments, sometimes performance trumps performance per watt. Sometimes, price/performance is the key metric. If Intel has been able to dramatically reduce the idle power draw of a server while not sacrificing clocks and therefore performance, this will give customers the best of both worlds. And it could give Intel and its server partners something that customers will pay a slight premium for. ®