NVIDIA: 20nm from TSMC won’t be cost efficient

As we close in on the theoretic limitations of silicon prices will increase exponentially. NVIDIA now says 20 nanometer at TSMC will hardly be cheaper than 28 nanometer, and the same goes for the successor at 14 nanometer.

The cost for new technologies and manufacturing in the semiconductor industry will only go from here. Earlier there was an idea of a breaking point, where we will have to reconsider if it is worthhile to move on to a new node, rather than trying to optimize for the one you are on. NVIDIA has openly commented on the issue and its manufacturing partner since many years, TSMC, gets harsh critique.

TSMC is the world’s largest foundry, which manufactures circuits for most companies that you can think of in the semiconductor industry. NVIDIA says that its partners will share wins and losses, and be open and honest when doing business. This sounds like NVIDIA is displeased with something that has been going on behind closed doors, but we get a better picture when look at the cost of manufacturing with future technologies.

Imagine a new graphics circuit, where the transistor price remains the same…

The transition to 28 nanometer is expected to be more price effective (price/transistor) than the 40 nanometer technology as yields improve. The transition to 20 nanometer, and the succeeding 14 nanometer technology is expected to bring marginal improvements in the price/transistor over the 28 nanometer technology. At 20 and 14 nanometer we should be able to fit about 2x and 4x as many transistors, on the same surface as the current 28 nanometer technology, which just reached the retail market. To NVIDIA and AMD that is in constant need of more transistors for their graphics circuits, it means that the manufacturing cost could become substantially higher when the price per transistor stays at the same level.

High wafer prices and lower yields will result in exponentially higher costs

The exponential increase in price there has been talk about shows quite clearly when you look at the price per wafer. Even if NVIDIA has not stated any exact prices the diagram is pretty clear. TSMC will have to work closer together with its partners in the future, and NVIDIA urges for a better cooperation to move over to 450 millimeter wafers soon. This would boost the area from the current 300 millimeter wafers with 125 percent, which would of course result in lower production costs in the long run. The problem is that the new wafer size demands whole new equipment in the fabs, which will require enormous investments from TSMC.

It is very rare for a company to openly critisize its manufacturing partners, and there are rumors that NVIDIA wants to work with GlobalFoundries in the future. The cost problems are not isolated to TSMC, but should apply to all actors in the semiconductor industry. At smaller nodes technologies like Double Patterning, EUV will become essential, while things like FinFET and SOI will also be required to keep leakage at an acceptable level, which in turn will drive up costs even further.

Also the design for actors like NVIDIA will become a lot more complicated and the validation periods will be much longer, which in turn leads to lower ROI (Return on Investment). In combination with lower yields it is a fact that newer manufacturing technologies will become more expensive for a foundry customer, and thus also the end user.

Also AMD have made similar statements, but it did not pinpoint or blame TSMC or GlobalFoundries. Newer technologies are not as attractive as before and AMD’s new VP Rory Read have made it clear for analysts that the company will not stress over to new nodes because of the costs. Together with a low capacity at TSMC’s 28 nm node the Deccan platform that was supposed to replace Brazos was skipped in favor of the coming Brazos 2.0.

It is rare for a company to openly critisize manufacturing partners like NVIDIA has done, but it shows that new manufacturing technologies will be anything but easy to develop and design for. The rule that smaller nodes will be cheaper will no longer apply, if you don’t have your own fabs and are selling enormous quantites, like Intel and Samsung. Intel has even opened its doors for others to use available capacities to make chips, which is just another way for it to keep the fabs working at maximum capacity.

Source: Extremetech