AMD shipped the Radeon HD 5870 in September and could not make enough of them, and the reason is not demand. The reason is that TSMC’s 40nm process, the one immersion lithography was supposed to make routine, is throwing yields nobody planned for. The node that was meant to be a victory lap turned into the first time the foundry model‘s customers learned what it feels like when the fab underneath them stumbles.
40nm is the shrink of TSMC’s 45nm, the first node where dry 193nm lithography finally gave out and immersion took over. You run the exposure through a film of purified water between the lens and the wafer, the higher refractive index buys you resolution, and on paper it is one more clean step down the curve. In practice immersion drags in a whole new defect catalogue: watermarks and bubbles and a chemistry at the water boundary the dry process never had to model. TSMC is learning that catalogue in production, on its customers’ wafers.
The specific failure is in the vias and contacts, the tiny vertical connections between metal layers, where 40nm geometries plus immersion patterning are throwing defects faster than the process can be tuned to suppress them. TSMC’s own guidance is to double up every via, two connections where one used to do, so a single defect cannot open the circuit. The customers who took that advice and whose dies are small, Altera and Qualcomm among them, are yielding fine. The ones running enormous dies are not. AMD’s GPU comes off the wafer working far less often than anyone budgeted for, and Nvidia’s even larger next-generation part is slipping into next year while the two companies argue about whose fault the yield is. A big die is a bigger target, and you cannot build a product launch around a number like the one TSMC’s defect density is producing right now.
The timing could not be worse, because this is the same year a second name showed up at the leading edge. AMD spun its factories out in March into a new company, GlobalFoundries, funded with Abu Dhabi money and openly aiming at TSMC’s customers, and in September it agreed to buy Chartered Semiconductor outright, vaulting in one move to the third-largest foundry in the world, Singapore fabs and all. TSMC still books around half of all foundry revenue, and for fifteen years the pure-play leading edge was effectively it and a gap behind. Now there is somebody else with real fabs and real scale and real ambition, arriving in the exact quarter TSMC’s yields hand every fabless customer a reason to ask what their second source looks like.
TSMC is also carrying 40nm with an old-style gate. The transistor still uses a polysilicon gate over a silicon-oxynitride dielectric, the same basic recipe stretched thinner every node, while Intel has been shipping a high-k dielectric and a metal gate since its 45nm parts in 2007. That gap is real and it shows up as leakage, the current a transistor wastes sitting still, which is exactly the wrong thing to waste in the battery-powered parts that are starting to dominate the order book. TSMC’s answer is to hold high-k metal gate for 28nm and do it gate-last, building the metal gate after the high-temperature steps instead of before. The harder way to do it, and the company is betting, the more manufacturable one.
The demand mix is shifting underneath all of this. The volume that used to come from PC graphics and networking now competes with a wave of phone chips, Qualcomm and the rest, riding a smartphone market that did not exist in this shape three years ago. Those parts care about leakage and power more than raw clock speed, which makes the 40nm leakage gap sting in precisely the segment growing fastest. The foundry that wins the next five years is the one that gets low-power right, and 40nm is not TSMC’s finest argument that it will be them.
The yields will recover. They always do. Morris Chang took the chief executive title back in June, at seventy-seven, and a founder does not return to babysit a node that is going fine; TSMC will grind the defect density down over the next few quarters and 40nm will end its life as a perfectly good node nobody remembers being angry about. What will not un-happen is the lesson. Every fabless company that spent this year explaining to its own customers why it could not ship learned that renting the best fab in the world means inheriting its bad quarters too, with none of the control and all of the blame. GlobalFoundries just turned that lesson into a sales pitch. TSMC spent a decade making the single-source bet look free. 40nm is the invoice.