AMD Will Use New GlobalFoundries 12nm Node for Future CPUs – GPUs At the GlobalFoundries Technology Conference this week, AMD’s CTO, Mark Papermaster, announced his firm would use GlobalFoundries’ new 12nm LP process node for upcoming products. GlobalFoundries said 12nm LP would begin production in 2018 and ramp up to volume fairly quickly.
The reason the company can do that, almost without question, is 12nm isn’t an actual process node by the old definition at all. It’s the kind of iterative improvement that foundries used to roll out as a half-node, or referred to as a new type of capability on the same process node.
TSMC had multiple flavors of 28nm, if you recall, each with its own density metrics and performance characteristics. As node shrinks become more difficult and take longer, foundries have risen to the challenge by creating a blizzard of PR and marketing terms to paper over the fact that they can’t deliver the performance improvements they used to. It’s hilarious, in a way, watching everyone bend over backwards to claim rigorous amounts of improvement when the actual results they can ship are less impressive relative to their predecessors every single year.
Image by Patrick Moorhead
AMD will deploy 12nm LP for both Zen CPUs and Vega GPUs in 2018, presumably including the Zen+ CPU that previously appeared on its roadmap. Part of the reason we know that GF’s 12nm LP is a modest adjustment to existing processes is that the company has been quite clear about its roadmap. GF has already said that it won’t deploy a 10nm solution, preferring instead to leapfrog for 7nm. Whether this 12nm process is a 14nm node with a few 7nm-class tweaks or just an improved 14nm ultimately doesn’t matter; the 15 percent density and 10 percent performance improvements that GlobalFoundries is claiming is right in line with similar claims by TSMC and Samsung for their improvements.
Foundry roadmaps are the closest thing the industry has to being set in stone, because deploying new nodes costs huge amounts of money. The nodes themselves take a substantial amount of time to build and complete, particularly when equipment changes are needed, and we already know that GF’s 7nm lines will be designed for EUV integration when available. Whether this is a purely marketing-driven maneuver or a necessary stopgap due to 7nm problems isn’t clear. But anyone expecting anything but the most modest of performance or power improvements is probably overselling the boost.
That’s not to say AMD won’t get any utility out of the new node — Vega could use a power consumption drop, while a few hundred MHz of additional CPU performance would give Ryzen a bit of a kick in lightly threaded or single-threaded scenarios, where it tends to be weakest against Intel. But absent any glaring chip design issues that have limited clocks or prevented lower power consumption, we can’t reasonably expect a 10 to 15 percent improvement in either metric to reinvent the wheel.