Google Servers Run with AMD and/or Intel Microprocessors?
Google Inc. is a darling of both the Wall Street and Silicon Valley communities where some of the best technical minds of a generation have been using their technological prowess to drive first search technology and then internet advertising revenues to heights not previously seen. Early on, Google began to build out their now considerable infrastructure with systems from "from service provider server all-star Rackable Systems." Later, Google began to design its own servers in-house with a great deal of emphasis on buying lower-cost, slower-clocked microprocessors that consume less energy in the hopes of reducing its power bill and keeping servers more reliably up and running.
As Google has grown over the last few years, their technological infrastructure has expanded considerably. In addition to acquiring thousands of microprocessors and tons of communication gear, "Silicon Valley gossip places Google right behind the Tier 1 server vendors as the largest purchaser of disks and memory." Google also apparently controls more network fiber than any other organization.
Google is notorious for its secrecy and has said little about which companies supply the microprocessors that go into its tremendous numbers of servers. That changed on Friday, January 26th 2007 when Google spokesman Barry Schnitt said, "We bought a small number of chips from Intel recently, but we continue to be supplied by more than one vendor." This terse Google statement apparently was made in response to rumors earlier last week that Intel had displaced AMD in its four quarter run as Google's chip supplier.
There is conflicting information about just who now is the microprocessor supplier of choice to Google. The Register reported on Sunday, January 28th 2007 in the article Google did not boot AMD for Intel that "Rumors of a full-on switch from AMD to Intel chips have forced Google to emerge from its warm, creepy secrecy cocoon. Contrary to claims from Intel, Google has not experienced Xeonmania and elevated Intel to favored supplier status. Rather, it's just got a few Xeons laying around its data centers."
On the other hand, Cnet.com on Monday, January 29th 2007 reported in an article entitled Intel inside again for new Google servers that "Intel, armed with a custom-designed motherboard, has reclaimed Google as a server customer after a year watching the search powerhouse give its business to Advanced Micro Devices… Google has begun buying Intel server components in high volume, said Pat Gelsinger, a co-general manager of Intel's Digital Enterprise Group, speaking about the Google relationship on an internal Intel blog entry Wednesday seen by CNET News.com. 'We're in business with the volume systems ramp under way,' he said."
Apparently this Google dust-up was caused by a posting Pat Gelsinger made on an internal Intel blog. As the Cnet.com story relates,
On his blog, Gelsinger said Intel had to create custom equipment to win back the business.
Google "went to the competitor's platform for the last four quarters of deployments, largely on the (operational cost) model they use to judge their purchases," Gelsinger said, a model that takes into account power and cooling costs, server performance, memory costs and other factors. Intel design teams "have been maniacal as we designed a unique board for them, developing a unique memory module with them, working every angle of the cost equation and engaging with our sales teams to get the business," Gelsinger said.
The Google change is emblematic of Intel's rising fortunes. Beginning in 2003, AMD capitalized on performance and power efficiency advantages of its Opteron processors to make its way into the server lines of the four top-tier server companies: IBM, Hewlett-Packard, Dell and Sun Microsystems. But in late 2006, Intel's new dual-core Xeon 5100 "Woodcrest" and quad-core Xeon 5300 "Clovertown" processors helped Intel reclaim some server chip market share and put major price pressure on AMD.
AMD expects that its fortunes will be boosted by its upcoming quad-core Barcelona processor, which the company promises will provide 40 percent better performance overall compared with Clovertown. Barcelona is due in mid-2007 and will plug into existing servers with dual-core Opteron chips.
Server power consumption has become a huge problem as companies grapple with increasingly power-hungry servers and rising electricity and cooling costs. Google is among those to sound the alarm about increasing power and cooling expenses.
Toomre Capital Markets LLC provides advisory services to Advanced Micro Devices and is quite impressed by the "performance per watt" results that the AMD Opteron servers provide to financial services firms. It certainly will be interesting to see if the upcoming quad-core Barcelona microprocessor delivers on its promised 40% improvement. If so, in the high-performance computing sector of the server market, the AMD "Barcelona" chips will do well against Intel's quad-core Xeon 5300 "Clovertown" processors. What remains unknown is just how much impact and how quickly the introduction of new hafnium alloys as insulator material and the move to 45 nanometer gate technology will have on future chip releases. (See previous TCM post here.)
However, what is clear is that complex financial calculations will be able to be done yet faster. This means that existing derivative models with Monte Carlo simulations will produce results sooner. Complex algorithmic trading programs (like selling Portfolio A to purchase Portfolio B while minimizing Implementation Shortfall) can be contemplated. The Future of Real-time Enterprise Risk Management will become ever more possible (and the de facto reality) for leading financial services firms.