Never Assume: AMD Sticking With GF At 28nm
Oops. Yesterday we published a story (since taken down) noting that AMD had apparently flip-flopped on its 28nm plans and would use TSMC for 28nm production in the 2012 timeframe rather than going with GlobalFoundries. The original piece we discussed ran Cens.com, which bills itself as "Taiwan's Economic News." We're not going to link it again, but the piece is still live if you care to read it. We heard from both AMD and GlobalFoundries a few hours after we published our original story; both companies categorically deny that there's been a change of plans with regard to AMD's contract with TSMC.
Although the original story was presented as fact rather than rumor, we should've realized there was something wrong from the very start. As we've previously discussed, there are two competing methods for building chips at 32nm and below: gate-first or gate-last.. It's unclear which approach will prove better in the long term; Intel and TSMC have opted to go for the more-difficult gate-last technology while GlobalFoundries went with gate-first for its 32nm and 28nm nodes.
It's no simple task to port a CPU from one foundry to another, but it's possible. That's not the case with gate-first / gate-last designs; a chip designed for fabrication using gate-first has to be re-designed to use gate-last (and vice versa). This should've been a clue that AMD wouldn't swap foundry partners. We apologize for any confusion this may have caused; AMD remains a major TSMC partner at 40nm, but Cern's claim that Sunnyvale will opt to build 28nm parts at TSMC is untrue.