We’ve discussed AMD’s 7nm plans several times this year, including the company’s decision to source its 7nm silicon for Vega’s machine learning iteration from TSMC. As Extreme Ultraviolet Lithography (EUV) ramps up and begins to enter production, a number of foundries including TSMC, Samsung, GlobalFoundries, and Intel are pursuing their own strategies and introduction plans. One interesting tidbit in a recent report sheds a bit more light on AMD’s 7nm plans and suggests we’ll see the company pursuing dual-source strategy well into next year.
In an article at EETimes, GlobalFoundries admits using similar pitches and SRAM cells to TSMC’s 7nm, precisely because it gives AMD more flexibility. The company’s chief technologist, Gary Patton, told EET’s Rick Merritt that AMD “will have more demand than we have capacity, so I have no issues with that.” The same article also confirmed that AMD will provide its first 7nm chip with tapeout expected later this year. IBM hardware and other ASICs will follow in 2019.
When I toured GlobalFoundries earlier this year, the company spoke to the difference between its strategy for 7nm FinFET designs, where EUV is being introduced first, and its plans for fully depleted SOI (FD-SOI). Right now, GF offers fully depleted SOI at 22nm, with 12nm ramping up into full production. GF, however, made it clear that it’s transitioning to FinFET technology nodes as a fast follower, while its focusing on building up its 22FDX and 12FDX as an industry differentiator — a unique capability that GF has and other companies don’t.
While a full discussion of the differences between FD-SOI and FinFET is beyond our scope here, FD-SOI has some capabilities that make it attractive to analog and RF circuits, which don’t scale well on FinFET designs. It also offers much lower design costs (wafer costs estimates and comparisons vary) and, in the case of GF’s 12FDX, can burst up to providing FinFET-equivalent performance over short periods of time. It’s potentially favored for low-power IoT devices and automotive applications, and given the strong competition GF faces from TSMC and Samsung, it makes sense that the company would be looking for an alternative way to stand out.
But it’s also interesting to see GF moving to align its 7nm roadmap in a way that makes it easier for AMD to design at both TSMC and GF, because that’s precisely what foundries typically don’t do. With a few abortive exceptions, like the Common Foundry Platform (now defunct), a chip built at one foundry simply can’t be built at another without a full redesign. This is why companies don’t typically move production from one firm to another, even in the face of substantial delays. The implication here is that while AMD and GF are no longer formally attached or affiliated with each other, AMD remains one of GF’s largest and most important customers.
The optics of GF’s decision to fast-follow at 7nm FinFET are clouded by the fact that mainstream CPU and GPU designs aren’t the major drivers of new nodes anymore. Ten years ago, it was common to see AMD and Nvidia GPUs as some of the first cutting-edge node shipments when TSMC had a new process to play with. Mobile hardware now dominates these early shifts. It’s not clear which 7nm hardware will ship first, but mobile SoCs, not GPUs or CPUs, have been the pipecleaners of choice since 20nm.
The implications of Patton’s comment, generally speaking, is that AMD will return to split 7nm manufacturing between TSMC and GF, with its APUs and CPUs likely built at one company and its GPUs at another. Between the two, it may make more sense to build its 7nm GPUs at TSMC, if that’s the company also handling Nvidia’s. While its CPU competition with Intel remains critical to its overall bottom line, AMD also knows that Intel’s 10nm process is delayed. AMD has always preferred GlobalFoundries for its CPUs — it moved Krishna and Wichita to TSMC only after the GF design failed — and we suspect the company will continue to use GF for CPUs and APUs at 7nm and below.