Ever since AMD announced it would build an ARM processor, there've been questions about how the company would balance its x86 and ARM obligations and product lines. Today, the company gave a major update to how it sees the future of these products. This year, AMD will debut its first Cortex-A57 based server (codenamed Seattle
). In 2015, AMD will launch a 20nm SoC family around the new Puma+ core and a second-gen Cortex-A57 chip -- and those two CPUs will be drop-in compatible with each other (codenamed Project Skybridge). In 2016, AMD will follow up with its own custom ARM
One of the major changes coming next year is that both the ARM and x86 chips will apparently be HSA-compatible. Right now, AMD's Puma+ SoC on 28nm isn't
HSA-capable -- these changes imply that AMD will do some major fabric upgrades in-between the 28nm and 20nm dies to integrate HSA functions and standardize the GPU IP across both chips. AMD didn't specify if it would upgrade the GPU at the heart of the architecture, but we're guessing the company will at least bump the feature set from GCN 1.0 to GCN 1.1 along with the HSA upgrade and the ARM socket compatibility. The two chips likely share a memory controller and southbridge as well, and the 20nm bump will give both the ARM and x86 cores improved power consumption and performance-per-watt over current gen products.
These announcements are a major shift in strategy for AMD, which has always pursued an x86-centric strategy. Up until now, it looked like the move towards ARM servers might be a side business confined to the lucrative server market -- an edge play to see if AMD could gain market share in a hot area, but not a sustained focus for the company's business. That's now off the table; AMD is clearly committing to a combined x86 / ARM roadmap across the entire product segment.
As for the 16nm ARM core, codenamed K12, AMD was cagey about its benefits or capabilities beyond noting that it's a clean design based on the ARM architecture. As we discussed a year ago, most companies that license ARM cores actually license the physical design for a Cortex-A9, A7, or A15, but the company does
offer an alternative architecture license. Only a few companies have historically gone this route (Qualcomm and Apple are two of the biggest), so AMD is entering a new field by competing with an architectural implementation of its own making. The company's custom architecture will be ready by 2016, and will likely debut on either TSMC's 16nm FinFET or GlobalFoundries' 14nm process nodes.
Finally, there's word that AMD has also built a new x86 architecture under Jim Keller. Keller stated that the new architecture will be "from scratch," but gave no other details on its design. A new microarchitecture could still leverage lessons learned from the Jaguar
products without necessarily inheriting IP from either part, though we'd be surprised if the chip was a completely
new design -- history has typically demonstrated that evolving older architectures is typically a better path than inventing new ones out of cloth. Then again, after Bulldozer, AMD may have decided it needed a fundamentally new approach.
Interestingly, while Project Skybridge is advertising x86/ARM SoC pin compatibility in the same socket, there's no word on whether the 2016 core will continue that trend. Presumably it will, unless AMD has turned to other methods for ensuring that x86 and ARM code can run on the same silicon.