Troubling News: AMD Moves 7nm GPU Production Back to TSMC

AMD disclosed its Q1 2018 earnings figures Wednesday, with remarkably strong sales that bucked seasonal trends and emphasized that Ryzen is driving AMD’s newfound profitability. During the conference call, Lisa Su announced that 7nm GPU production was on track, and that AMD still expects to ship a 7nm Vega in the machine learning market later this year. That part, as we’ve reported before, is expected to be a “pipecleaner” — a design used to ensure that a company’s new node is ready for prime time before mass market parts are run through it.

According to CEO Lisa Su, however, that chip isn’t being built at GlobalFoundries — or, at least, it isn’t being built exclusively at GlobalFoundries. Here’s what she said:

So our foundry strategy is to use both TSMC and GlobalFoundries on the first 7-nanometer product. We are using TSMC for that product and we have a very strong relationship with them. And so, we do see a good momentum on it from what we see, and I’m not concerned about capacity.

That’s not a very good sign. Back when AMD first spun off its foundries and created GlobalFoundries, it signed an agreement stating that it would move all of its 28nm CPU and GPU production over to the new foundry. What happened next was a disaster for both companies. GlobalFoundries struggled to get yield on Llano, Bulldozer tanked, and AMD had to cancel its original plans to launch low-power APUs at GF (Krishna, Wichita) and move those chips to TSMC. AMD’s 28nm GPUs were never built at GlobalFoundries and AMD paid GF a substantial premium for failing to live up to its original agreement. GF’s own custom 14nm process, 14XM, was later cancelled, and GlobalFoundries became a second-source fab for Samsung.

AMD’s Ryzen 7 2700X is proof that GlobalFoundries’s licensed 14nm process is up and running smoothly, but there’ve always been questions about whether the foundry could deliver 7nm process technology on time. Furthermore, it makes little sense to bring up a chip that’s supposed to sell into a relatively small market (and machine learning is a small market right now) at two foundries at once.

Under the terms of its original agreement with GF — an agreement that’s still in effect to the best of our knowledge — AMD will have to pay GF a penalty for continuing to use TSMC. That alone is enough to raise questions about why AMD is building Vega with its old foundry partner, and it’s not the only disquieting fact to surface in the past few months. A few weeks after we toured GlobalFoundries, the company announced that its CEO, Sanjay Jha, was stepping down to be replaced by Thomas Caulfield.

GlobalFoundries is now on its fourth CEO in a decade. If we put the situation in terms of Roman Emperors, that’s far better than the Year of the Four Emperors (self-explanatory), and roughly on par with the average length of rule during the Crisis of the Third Century. (Note: From this point forth, I intend to suggest that all tech companies be compared with obscure bits of Imperial Roman history). Then again, it was called the “Crisis” of the Third Century, as opposed to “That 51 Year Period in the Third Century When Everything Was Awesome.”

If I’m being honest, I’ve been worried about the 14/12nm – 7nm transition since Ryzen launched. Seeing GF’s EUV hardware being installed and the company’s 12nm and 14nm facility in production allayed some of those concerns. But the decision to replace Jha with Caulfield raised them again, and AMD’s decision to push some production back to TSMC doesn’t look great for GF’s 7nm progress.

To be clear, this is speculation on my part. It could be that GF’s 7nm progress is fine, but TSMC’s timetable is better aligned with AMD’s product launch. It could be that GF is having trouble with 7nm GPU tech, but foresees no issues on the CPU side of the equation. It could be that GF is having 7nm trouble, but that they’ll have the kinks worked out by the time Ryzen 2 is ready for launch. And, of course, it’s possible that AMD is only test-piloting 7nm Vega at TSMC in a limited run and that mainstream GPU production will remain anchored to the GF foundry in Malta. Heck, it’s even possible that the old agreement that required AMD to move all production to GF has been waived or amended to give AMD more flexibility. But based on what we know right now, today, I’m not sure this is good news.