May 16, 2014, 2:53 AM — For quite some time now, AMD’s CPU strategy seems to have been “good enough.” When it came out with new chips, they were often around the middle to upper-middle of what Intel already had on the market, which mean when Intel shipped something new, it left AMD behind.
The result was an AMD that never really grew. It hung on to what it had but that was about it. It was very good at making affordable chips, and AMD-powered PCs could save consumers $100 or more over comparable Intel machines.
But the real fight of a decade ago, when AMD was first to 1GHz, the first to 64-bit, the first to dual core, seemed missing. It’s not surprising since the company was facing a real threat to its survival. But with a gravy train from the gaming consoles, it looks like the company is ready for a fresh battle, with a familiar face at the helm.
A report on a Chinese tech site Expreview.com says Jim Keller, one of the company’s main designers who left the company years ago, only to return from Apple in 2012, is leading the design of a new architecture for 2016. This would be separate from the K12/Skybridge x86/ARM hybrid core that was recently announced.
Keller’s work is spectacular, putting him on par with AMD’s Dirk Meyer and Intel’s Pat Gelsinger. He worked with Meyer on the DEC Alpha 64-bit RISC processor, then came to AMD to work on what would be the K7 architecture, more commonly known as the Athlon. He was also involved in the development of the Hyper Transport interface and 64-bit x86 specs, which appeared in the K8 design.
He left AMD for a company called Sibyte, later acquired by Broadcom, and then joined PA Semi as the vice president of engineering in 2004. Apple acquired PA Semi in 2008, and Keller went to Apple, where he worked on the A4 and A5 SoCs.
His task will be a new microarchitecture to overcome some of the shortcomings in AMD’s current generation microarchitecture, called Bulldozer. Bulldozer adopted clustered multi-thread (CMT) designs used in the Alpha 21264 microprocessor, which isn’t surprising given how much overlap there has been between the Alpha and Athlon.
The processor runs one or more dual-core modules that communicate via HyperTransport if there is more than one module (two for quad-core, etc). The problem is that each module has two integer cores but one floating point unit, which is shared. And given how much computing is floating point calculations to begin with, this means a Bulldozer is inherently less efficient than Intel’s chips, which have the FPU in every core, and Intel’s FPU is ridiculously fast, too.
What Keller will do, no one knows. And AMD would be nuts to tip its hand. The most logical move for Keller would be to dump the CMT design in favor of a design with simultaneous multi-threading (SMT), which is what Intel does (and IBM’s Power and Oracle’s Sparc line). AMD for a long time eschewed hyperthreading in favor of cores; its Opteron server chips come in 12- and 16-core designs, for example.
So if he throws out CMT, that basically means a whole new design from the ground up with much more focus on performance. Expreview rightly points out that the Athlon FX and Opteron line don’t have a roadmap beyond 2015, and Keller’s project is set for a late 2015 introduction and ship in 2016.
Bulldozer is getting a little long in the tooth. It’s gone through three revisions: Piledriver, Steamroller, and Excavator. Yes, AMD likes its heavy equipment. Chips with the Excavator cores will appear this year under the codenames Carrizo and Toronto.