Even the companies that have fabs use foundries. It's an easier and quicker way to increase capacity and allows them to decrease capacity without running their own fabs at less than full capacity. It's a nice hypothesis, but the actual results aren't that easy. For a fabless company to build their own fab is a huge task. The time and cost up front is huge. I don't see that happening for Qualcomm.
The only "jump" at Qualcomm is the suicide that happened last month on June 11. Well another one 2 years ago, but who's really counting now right? It's no mystery that the June 11 incident didn't get reported in ANY news outlets, in any sort of media. When there's a suicide in the city, it gets reported. Question is how much money did Qualcomm pay off to keep this a hush hush, just so that their stocks don't get a dive? Very unethical of a company of that size. The fact that everything's so secretive makes employees working there want to spread the word out even more.
I think that if QCT had their own fab, they'd enjoy the same advantage that Intel does - that is, optimizing your process for one design, rather than just using an SRAM or something and then having to tweak your devices for each customer that comes in after development is done. But your larger point is right - management motives are usually less than transparent.
I agree, it'd be crazy for Qualcomm to jump into a bleeding edge logic fab - but hey, it was crazy for ATIC to create Global Foundries and jump into the foundry business with no pre-existing expertise. But they did it anyhow, and it seems like they're able to tread water. So maybe Qualcomm will pop a crazy pill one morning and take a go at it.
Those who were paying attention when DFM had its 15 minutes of fame at DAC and VCs should have seen that the bandaid solutions which worked at 65 and 45nm wouldn't work at sub-20nm IF EUV didn't make it. And those who remember that EUV is still x-ray should have seen that EUV wouldn't make it. The only way to build 10nm logic devices will be with completely seamless integration of design and manufacturing. For Intel and other such IDMs, this is second nature. Fabless/foundry will require a new model. Perhaps reintegration? Buying a fab? Renting a fab line? Otherwise, IDMs rule.
Well, it may figure that a fabless company, even Qualcomm, may be averse to facing the fab-intensive risk and complexities of scaling to the next node, to the point of not wanting to face it earlier than necessary. So maybe it is not the 'fabless model' that is the trouble, but the fabless mindset.
How would yield be better if you own the fab? Your own learning curve would be at best at par or worse than TSMC at each nodes. If you do the math right, it does not even make sense for a company like Qualcomm to own a fab. They always had a option to purchase additional capacity ahead of time with take or pay option in their supply agreement if they planned it right. The problem is that they never execised its option until it's too late. To me it's a Qualcomm's indecision and management issue and not able to plan their demand picture correctly. They're simply shifting investors' focus from inability to manage their demand to supply issue. There is no supply issue here if they planned it ahead. Now they are going around every foundries and trying to drum up the capacity - which in turn is a wrong move.
The issue of pre-booked capacity is also inextricably linked to yield. Which then begs the question as to whether low yields are the due to poor design (the fabless) or the process (the foundry).
Qualcomm may be arguing that at 80 percent yield they would have had enough wafers but at 20 percent yield they only have one quarter of what they need.
TSMC's response may well be that at this stage in a roll out lower yields are quite normal and that Qualcomm should either have bought more wafers...or should make its customers wait for their chips.
As such, mentioning alternative manufacturing strategies may be a way of putting pressure on to TSMC...a coded way of saying "You are my supplier and are meant to be trying to keep me happy. I am not happy and have to think about my future here."
Blog Doing Math in FPGAs Tom Burke 13 comments For a recent project, I explored doing "real" (that is, non-integer) math on a Spartan 3 FPGA. FPGAs, by their nature, do integer math. That is, there's no floating-point ...