Recently, the performance gap between Intel and AMD has been closing -- first with AMD leading and gaining some much-earned credibility, then, more recently, with Intel making a comeback. Intel also apparently leads in performance per watt now, too.
But Google had to know about the "comeback" when they announced that they were going to use AMD Opterons in the future a little while ago.
Mark Edelstone, a prominent equity analyst writes:
"Based on our various research efforts, we believe that most of Google's near-term server purchases will use AMD's Opteron for the first time."
But why? If the CEO of Intel, Paul Otellini, sits on Google's board of directors, Google would have known what was up Intel's sleeve. Paul, I am sure, made a persuasive case for Intel. Google using AMD products would be an unofficial endorsement, and a major PR gain for AMD.
Why? Because AMD has this new technology that allows you to hook in specialized chips to accelerate applications via HyperTransport. Google could, for example, design a chip that astronomically speeds query parsing, or perhaps do something as obvious as accelerating Java applications. That, by the way, would have the "viral" effect of increasing the execution speed of every Java application they employ. Google is known to use Java pervasively.
These little toys could work wonders for execution speed, and give Google a proprietary edge. Intel cannot deliver the goods in that regard.
That's why Google probably chose AMD, much to Paul's chagrin.
July 25th, 2006 at 10:03 pm
[...] Since they must have known that Intel would regain the speed crown this year, the question becomes “Why?”. Jaimie Sirovich suggests: Why? Because AMD has this new technology that allows you to hook in specialized chips to accelerate applications via HyperTransport. Google could, for example, design a chip that astronomically speeds query parsing, or perhaps do something as obvious as accelerating Java applications. That, by the way, would have the “viral” effect of increasing the execution speed of every Java application they employ. Google is known to use Java pervasively. [...]
August 25th, 2006 at 5:24 am
Take a look at this google TechTalk video from february 2006:
http://video.google.com/videoplay?docid=-4969729965240981475
It give some very interesting hindsight about what they could be thinking of doing.
If they do something like:
1) Analysing their most CPU intensive algorithms/subroutines
2) Create a team of Ph D. specialized in FPGA
3) Use the team to enhance their most intensive algorithms
If they want to simply/automate things even more, they could create some functions in a lib or add them to their Map/Reduce library. This would allow easy distribution of FPGA configuration files for the FPGA co-processors.
Another interesting paper is come from noone else than microsoft R&D:
http://research.microsoft.com/users/mbudiu/power-efficiency04.ppt