Samsung, GlobalFoundries to fab next-gen AMD GPUs, APUs

A new report is claiming Samsung and GlobalFoundries will team up to build next-generation AMD APUs and GPUs at the 14nm node. In the past, AMD has sourced its big-core APUs from GlobalFoundries and its smaller, budget CPU cores, graphics cards, and game consoles from TSMC. Shifting more of this business to the Samsung / GlobalFoundries match-up would be a significant shift from AMD's previous strategy.

Then again, it's arguably more of what AMD originally intended to do with the GlobalFoundries spin-off in the first place. When AMD created GF, it signed wafer agreements that allowed it to continue producing GPU hardware at TSMC, but stipulated that all future 28nm products would be built at GF. This never happened, thanks to GlobalFoundries' problems with Krishna and Wichita back several years ago. We had it on good authority that 20nm versions of the modern consoles were also planned for GlobalFoundries, before that node was canceled.

We can't talk about AMD's upcoming GPUs just yet, but the Korean Times is claiming(Opens in a new window) that the upcoming high-end APU, Greenland, will offer up to 2x higher performance-per-watt than Fiji. That's entirely reasonable, given that all next-generation lithography nodes offer companies a range of performance and power consumption to target. The figures we've seen for 14nm in general have suggested either 50% reduced power at the same performance or 20-30% improved performance at equivalent power.

AMD is building Greenland and its other GPUs on Samsung's 14nm LPP (Low Power Plus) process, but GF has hinted to us that it's done some customization to make sure the process node is suited to higher power GPUs. We don't know yet if AMD has split its foundry orders between TSMC and GF for next-gen graphics or not -- it's entirely possible that the company will do high-end desktop cards, with 150W+ TDPs at TSMC, and mobile hardware at GF.

AMD would likely prefer to avoid such scenarios, since they increase overhead costs associated with bringing up multiple designs and different foundries. But the company plays coy on its foundry partners and doesn't like to disclose who builds which parts until we can read the labels for ourselves.