In a shocking but not totally unexpected development, Advanced Micro Devices announced Oct. 29 that it would license a future 64-bit core from ARM and develop it for use as a microprocessor for the data center.
Instead of taking a license to the ARM architecture itself, AMD will license a single, undisclosed 64-bit ARM core and surround it with its existing intellectual property, such as I/O. The exact core will be disclosed by ARM Oct. 30, when the company will make the announcement at its developer conference in Silicon Valley.
“I think that this is a key moment in the industry,” said Rory Read, the chief executive of AMD, at an Oct. 29 press conference in San Francisco. “I don’t think that there is any doubt about it. Today is about what AMD is about at its core, and it’s in our DNA to innovate, and to think about what’s next… we look to disrupt the status quo, we look to drive the industry to where it needs to go.”
In broad strokes, the agreement means that AMD will develop an ARM based Opteron. AMD currently develops Opterons using the same X86-based technology used by Intel, although with a few proprietary techniques. With the 64-bit ARM core, AMD will do the same, but replace the higher-performance Opteron chip with the optimized ARM core that David Feldman, general manager of the Data Center Solutions Division at AMD, promised could save about a third to half the power of its standard Opteron processors.
The key piece of intellectual property will be the Freedom Fabric, a fabric chip that AMD acquired with the acquisition of SeaMicro. AMD unveiled the SM15000 server a short time ago, using a Freedom Fabric controller that can either be attached to an Intel Xeon or an AMD Opteron. Now, the technology will connect to the ARM architecture as well.
ARM and AMD already have a previous relationship. In June, AMD said it would integrate the ARM TrustZone security solution into its APUs (Accelerated Processing Units), which integrate the processor and graphics logic into a single component. But a more formal partnership has been suspected (and pushed for, by some Wall Street analysts) for months.
The question, of course, is the potential significance of ARM architecture in the data center. Analyst firms such as Gartner have termed the potential of low-power servers “minuscule.” Nonetheless, AMD and ARM convinced Dell and Facebook to join them onstage, and both endorsed the future of ARM servers.
Read, in response to a question from SlashDataCenter, said the potential for low-power serves in the data center was “well into a double-digit percentage” in three to five years.
Feldman said in an interview that he expects ARM to hand off the design to AMD in early 2013; then AMD will work to test, verify, and ultimately manufacture the design. That will take roughly about a year, he said, meaning that AMD is scheduled to deliver the parts in about 2015.
According to Lisa Su, the senior vice president of AMD’s global business units, AMD will ship the first ARM-based Opterons in 2014.
Analysts said that, at least for a position in the data center, the partnership is a win-win for both sides.
“The significance is that ARM covets a position in the data center, and AMD has that position—yes, it’s a bit tenuous, but it’s there,” said Nathan Brookwood, an analyst with Insight 64. “And with the OEM contacts that AMD can bring to the table, it’s a whole lot better for ARM to have an ally like AMD, than annually like Calxeda.”
The agreement doesn’t preclude AMD from taking an architectural license. But Brookwood said more is at stake than just a unique design.
“If ARM’s doing a decent job in designing a core and they have a lot of expertise in this, it takes an enormous amount of hubris to believe that you can design a better core than they can,” he added. “And certainly if you’re AMD, time-to-market matters. And so designing a new core over multiple years, where as taking an existing core and throwing it in with your Freedom Fabric and things, that’s an 18-month type of deal.”