If you know anything about semiconductors, you know two things, smaller geometry ICs are hard to make, and Nvidia is miserable at making them. The company had <2% yields on their first big 40nm chip, the initial sample run of Fermi/GF100, and SemiAccurate’s sources tell us that yields never broke 20% over the entire production life. It was hot, slow and unmanufacturable, but somehow this didn’t affect margins. They still made money on it and the related GF1xx derivatives even though yields were lower than the competition. It didn’t make sense.
Nvidia blamed everyone but themselves for the debacle, TSMC took the brunt of this on 40nm. Previously, it was customers, users, and anyone else they could point a microphone at. Prior to 28nm production, the problems that haunted Nvidia on prior generations were completely solved according to no less than the CEO Jen-Hsun Huang.
Oh boy did we laugh at that one, until we realized the press actually believed it, then we were a tad depressed. None of the problems SemiAccurate saw were addressed, much less dealt with. Needless to say, anyone who understands semiconductors could see there were still some acute and rather painful problems with Nvidia’s 28nm chips. The 28nm Fermi shrinks were respun, delayed, and have since vanished, something that does not sync with claims of a problem free 28nm ramp.
There were some obvious questions about the future, including Nvidia’s ability to fulfill some pretty big contracts, Apple notwithstanding. Needless to say, SemiAccurate’s crystal ball was a lot less muddy than Nvidia’s. Truth can be bent to a point, but after a while, it doesn’t make any sense to keep on bending, but this never stopped Nvidia. Luckily, the SEC rules make some things a little less flexible than others, so eventually something had to break. Then came the admissions, 28nm yields were horrible, so horrible that Jen-Hsun was forced to admit it to Wall Street.
According to Nvidia, the yield problems on 28nm seemed to affect everyone equally. SemiAccurate’s checks of other companies running 28nm product at TSMC did not come up with the same data points, something we can’t attribute to nuance variations in interpretation. Yield problems, contract problems, wafer shortages, missing product lines, and the abject inability to supply a decidedly low volume product on one side, plentiful volume from the competition making notably larger parts on the same process on the other. It is truly a puzzler. Well, not really.
That brought up the most obvious question, if 40nm yields were still bad, and 28nm was likely not economically viable yet for Nvidia, even if it is for AMD, how can they keep returning the numbers that they do? The single most relevant cost to any semiconductor company is yield, and Nvidia’s yields are not in line with their financial returns. Finally, a real puzzler.
Luckily for us, the answer could be found in the latest 10-K, filed on March 13, 2012, it confirms a long running rumor about Nvidia’s wafer pricing. Buried in the midst as if to obfuscate bad news was this gem, “Because of our potentially limited access to wafer foundry capacity and our recent transition to a wafer buy model where the costs of our products are based on the price per wafer versus price per functional die, decreases in manufacturing yields could result in an increase in our costs and force us to allocate our available product supply among our customers“. It looks like the differences between yield and financial results had a buffer for 40nm products, basically the last two years. It also looks like this buffer went away at precisely the wrong moment.
So that puzzler solved, we are faced with another. If 40nm was never really working well for Nvidia, but it was for everyone else, and 28nm is demonstrably worse for Nvidia, but not for everyone else, and the financial buffer is gone, what happens next? Low yields mean low margins, but also the need to run more expensive wafers in order to fulfill large volume contracts. If there are wafer shortages like Nvidia is actively warning about (see both their most recent 10-K and their most recent quarterly call) on top of this, what happens to large contracts? Since SemiAccurate is focused on technology, not finance, we will leave that up to the experts to fret over.S|A
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017